Thursday, March 26, 2015

PyPy 2.5.1 Released

PyPy 2.5.1 - Pineapple Bromeliad

We’re pleased to announce PyPy 2.5.1, Pineapple Bromeliad following on the heels of 2.5.0. You can download the PyPy 2.5.1 release here:
We would like to thank our donors for the continued support of the PyPy project, and for those who donate to our three sub-projects, as well as our volunteers and contributors. We’ve shown quite a bit of progress, but we’re slowly running out of funds. Please consider donating more, or even better convince your employer to donate, so we can finish those projects! The three sub-projects are:
  • Py3k (supporting Python 3.x): We have released a Python 3.2.5 compatible version we call PyPy3 2.4.0, and are working toward a Python 3.3 compatible version
  • STM (software transactional memory): We have released a first working version, and continue to try out new promising paths of achieving a fast multithreaded Python

  • NumPy which requires installation of our fork of upstream numpy, available on bitbucket
We would also like to encourage new people to join the project. PyPy has many layers and we need help with all of them: PyPy and Rpython documentation improvements, tweaking popular modules to run on pypy, or general help with making Rpython’s JIT even better.

What is PyPy?

PyPy is a very compliant Python interpreter, almost a drop-in replacement for CPython 2.7. It’s fast (pypy and cpython 2.7.x performance comparison) due to its integrated tracing JIT compiler.

This release supports x86 machines on most common operating systems (Linux 32/64, Mac OS X 64, Windows, and OpenBSD), as well as newer ARM hardware (ARMv6 or ARMv7, with VFPv3) running Linux.

While we support 32 bit python on Windows, work on the native Windows 64 bit python is still stalling, we would welcome a volunteer to handle that.


  • The past months have seen pypy mature and grow, as rpython becomes the goto solution for writing fast dynamic language interpreters. Our separation of Rpython from the python interpreter PyPy is now much clearer in the PyPy documentation and we now have seperate RPython documentation. Tell us what still isn’t clear, or even better help us improve the documentation.
  • We merged version 2.7.9 of python’s stdlib. From the python release notice:
    • The entirety of Python 3.4’s ssl module has been backported. See PEP 466 for justification.
    • HTTPS certificate validation using the system’s certificate store is now enabled by default. See PEP 476 for details.
    • SSLv3 has been disabled by default in httplib and its reverse dependencies due to the POODLE attack.
    • The ensurepip module has been backported, which provides the pip package manager in every Python 2.7 installation. See PEP 477.

  • The garbage collector now ignores parts of the stack which did not change since the last collection, another performance boost
  • errno and LastError are saved around cffi calls so things like pdb will not overwrite it
  • We continue to asymptotically approach a score of 7 times faster than cpython on our benchmark suite, we now rank 6.98 on latest runs
Please try it out and let us know what you think. We welcome success stories, experiments, or benchmarks, we know you are using PyPy, please tell us about it!
The PyPy Team

Friday, March 13, 2015

Pydgin: Using RPython to Generate Fast Instruction-Set Simulators

Note: This is a guest blog post by Derek Lockhart and Berkin Ilbeyi from Computer Systems Laboratory of Cornell University.

In this blog post I'd like to describe some recent work on using the RPython translation toolchain to generate fast instruction set simulators. Our open-source framework, Pydgin [a], provides a domain-specific language (DSL) embedded in Python for concisely describing instruction set architectures [b] and then uses these descriptions to generate fast, JIT-enabled simulators. Pydgin will be presented at the IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS) and in this post we provide a preview of that work. In addition, we discuss some additional progress updates that occurred after the publishing deadline and will not appear in the final paper [1].

Our area of research expertise is computer architecture, which is perhaps an unfamiliar topic for some readers of the PyPy blog. Below we provide some brief background on hardware simulation in the field of computer architecture, as well as some context as to why instruction set simulators in particular are such an important tool.

Simulators: Designing Hardware with Software

For computer architects in both academia and industry, a key step in designing new computational hardware (e.g., CPUs, GPUs, and mobile system-on-chips) is simulation [c] of the target system. While numerous models for simulation exist, three classes are particularly important in hardware design.

Functional Level models simulate the behavior of the target system. These models are useful for creating a "golden" reference which can serve as an executable specification or alternatively as an emulation platform for software development.

Cycle Level models aim to simulate both the behavior and the approximate timing of a hardware component. These models help computer architects explore design tradeoffs and quickly determine things like how big caches should be, how many functional units are needed to meet throughput targets, and how the addition of a custom accelerator block may impact total system performance.

Register-Transfer Level (RTL) models specify the behavior, timing, and resources (e.g., registers, wires, logic gates) of a hardware component. RTL models are bit-accurate hardware specifications typically written in a hardware description language (HDL) such as Verilog or VHDL. Once verified through extensive simulation, HDL specifications can be passed into synthesis and place-and-route tools to estimate area/energy/timing or to create FPGA or ASIC prototypes.

An instruction set simulator (ISS) is a special kind of functional-level model that simulates the behavior of a processor or system-on-chip (SOC). ISSs serve an important role in hardware design because they model the instruction set architecture (ISA) interface: the contractual boundary between hardware designers and software developers. ISSs allow hardware designers to quickly experiment with adding new processor instructions while also allowing software developers to build new compilers, libraries, and applications long before physical silicon is available.

Instruction-Set Simulators Must be Fast and Productive

Instruction-set simulators are more important than ever because the ISA boundary has become increasingly fluid. While Moore's law has continued to deliver larger numbers of transistors which computer architects can use to build increasingly complex chips, limits in Dennard scaling have restricted how these transistors can be used [d]. In more simple terms, thermal constraints (and energy constraints in mobile devices) have resulted in a growing interest in pervasive specialization: using custom accelerators to more efficiently perform compute intensive tasks. This is already a reality for designers of mobile SOCs who continually add new accelerator blocks and custom processor instructions in order to achieve higher performance with less energy consumption. ISSs are indispensable tools in this SOC design process for both hardware architects building the silicon and software engineers developing the software stack on top of it.

An instruction set simulator has two primary responsibilities: 1) accurately emulating the external execution behavior of the target, and 2) providing observability by accurately reproducing the target's internal state (e.g., register values, program counter, status flags) at each time step. However, other qualities critical to an effective ISS are simulation performance and designer productivity. Simulation performance is important because shorter simulation times allow developers to more quickly execute and verify large software applications. Designer productivity is important because it allows hardware architects to easily experiment with adding new instructions and estimate their impact on application performance.

To improve simulation performance, high-performance ISSs use dynamic binary translation (DBT) as a mechanism to translate frequently visited blocks of target instructions into optimized sequences of host instructions. To improve designer productivity, many design toolchains automatically generate ISSs from an architectural description language (ADL): a special domain-specific language for succinctly specifying instruction encodings and instruction semantics of an ISA. Very few existing systems have managed to encapsulate the design complexity of DBT engines such that high-performance, DBT-accelerated ISSs could be automatically generated from ADLs [e]. Unfortunately, tools which have done so are either proprietary software or leave much to be desired in terms of performance or productivity.

Why RPython?

Our research group learned of the RPython translation toolchain through our experiences with PyPy, which we had used in conjunction with our Python hardware modeling framework to achieve significant improvements in simulation performance [2]. We realized that the RPython translation toolchain could potentially be adapted to create fast instruction set simulators since the process of interpreting executables comprised of binary instructions shared many similarities with the process of interpreting bytecodes in a dynamic-language VM. In addition, we were inspired by PyPy's meta-tracing approach to JIT-optimizing VM design which effectively separates the process of specifying a language interpreter from the optimization machinery needed to achieve good performance.

Existing ADL-driven ISS generators have tended to use domain-specific languages that require custom parsers or verbose C-based syntax that distracts from the instruction specification. Creating an embedded-ADL within Python provides several benefits over these existing approaches including a gentler learning curve for new users, access to better debugging tools, and easier maintenance and extension by avoiding a custom parser. Additionally, we have found that the ability to directly execute Pydgin ISA descriptions in a standard Python interpreter such as CPython or PyPy significantly helps debugging and testing during initial ISA exploration. Python's concise, pseudocode-like syntax also manages to map quite closely to the pseudocode specifications provided by many ISA manuals [f].

The Pydgin embedded-ADL

Defining a new ISA in the Pydgin embedded-ADL requires four primary pieces of information: the architectural state (e.g. register file, program counter, control registers), the bit encodings of each instruction, the instruction fields, and the semantic definitions for each instruction. Pydgin aims to make this process as painless as possible by providing helper classes and functions where possible.

For example, below we provide a truncated example of the ARMv5 instruction encoding table. Pydgin maintains encodings of all instructions in a centralized encodings data structure for easy maintenance and quick lookup. The user-provided instruction names and bit encodings are used to automatically generate decoders for the simulator. Unlike many ADLs, Pydgin does not require that the user explicitly specify instruction types or mask bits for field matching because the Pydgin decoder generator can automatically infer decoder fields from the encoding table.

encodings = [
  ['adc',      'xxxx00x0101xxxxxxxxxxxxxxxxxxxxx'],
  ['add',      'xxxx00x0100xxxxxxxxxxxxxxxxxxxxx'],
  ['and',      'xxxx00x0000xxxxxxxxxxxxxxxxxxxxx'],
  ['b',        'xxxx1010xxxxxxxxxxxxxxxxxxxxxxxx'],
  ['bl',       'xxxx1011xxxxxxxxxxxxxxxxxxxxxxxx'],
  ['bic',      'xxxx00x1110xxxxxxxxxxxxxxxxxxxxx'],
  ['bkpt',     '111000010010xxxxxxxxxxxx0111xxxx'],
  ['blx1',     '1111101xxxxxxxxxxxxxxxxxxxxxxxxx'],
  ['blx2',     'xxxx00010010xxxxxxxxxxxx0011xxxx'],
  # ...
  ['teq',      'xxxx00x10011xxxxxxxxxxxxxxxxxxxx'],
  ['tst',      'xxxx00x10001xxxxxxxxxxxxxxxxxxxx'],

A major goal of Pydgin was ensuring instruction semantic definitions map to ISA manual specifications as much as possible. The code below shows one such definition for the ARMv5 add instruction. A user-defined Instruction class (not shown) specifies field names that can be used to conveniently access bit positions within an instruction (e.g. rd, rn, S). Additionally, users can choose to define their own helper functions, such as the condition_passed function, to create more concise syntax that better matches the ISA manual.

def execute_add( s, inst ):
  if condition_passed( s, inst.cond() ):
    a,   = s.rf[ inst.rn() ]
    b, _ = shifter_operand( s, inst )
    result = a + b
    s.rf[ inst.rd() ] = trim_32( result )

    if inst.S():
      if inst.rd() == 15:
        raise FatalError('Writing SPSR not implemented!')
      s.N = (result >> 31)&1
      s.Z = trim_32( result ) == 0
      s.C = carry_from( result )
      s.V = overflow_from_add( a, b, result )

    if inst.rd() == 15:

  s.rf[PC] = s.fetch_pc() + 4

Compared to the ARM ISA Reference manual shown below, the Pydgin instruction definition is a fairly close match. Pydgin's definitions could certainly be made more concise by using a custom DSL, however, this would lose many of the debugging benefits afforded to a well-supported language such as Python and additionally require using a custom parser that would likely need modification for each new ISA.

if ConditionPassed(cond) then
   Rd = Rn + shifter_operand
   if S == 1 and Rd == R15 then
     if CurrentModeHasSPSR() then CPSR = SPSR
   else UNPREDICTABLE else if S == 1 then
     N Flag = Rd[31]
     Z Flag = if Rd == 0 then 1 else 0
     C Flag = CarryFrom(Rn + shifter_operand)
     V Flag = OverflowFrom(Rn + shifter_operand)

Creating an ISS that can run real applications is a rather complex task, even for a bare metal simulator with no operating system such as Pydgin. Each system call in the C library must be properly implemented, and bootstrapping code must be provided to set up the program stack and architectural state. This is a very tedious and error prone process which Pydgin tries to encapsulate so that it remains as transparent to the end user as possible. In future versions of Pydgin we hope to make bootstrapping more painless and support a wider variety of C libraries.

Pydgin Performance

In order to achieve good simulation performance from Pydgin ISSs, significant work went into adding appropriate JIT annotations to the Pydgin library components. These optimization hints, which allow the JIT generated by the RPython translation toolchain to produce more efficient code, have been specifically selected for the unique properties of ISSs. For the sake of brevity, we do not talk about the exact optimizations here but a detailed discussion can be found in the ISPASS paper [1]. In the paper we evaluate two ISSs, one for a simplified MIPS ISA and another for the ARMv5 ISA, whereas below we only discuss results for the ARMv5 ISS.

The performance of Pydgin-generated ARMv5 ISSs were compared against several reference ISSs: the gem5 ARM atomic simulator (gem5), interpretive and JIT-enabled versions of SimIt-ARM (simit-nojit and simit-jit), and QEMU. Atomic models from the gem5 simulator were chosen for comparison due their wide usage amongst computer architects [g]. SimIt-ARM was selected because it is currently the highest performance ADL-generated DBT-ISS publicly available. QEMU has long been held as the gold-standard for DBT simulators due to its extremely high performance, however, QEMU is generally intended for usage as an emulator rather than a simulator [c] and therefore achieves its excellent performance at the cost of observability. Unlike QEMU, all other simulators in our study faithfully track architectural state at an instruction level rather than block level. Pydgin ISSs were generated with and without JITs using the RPython translation toolchain in order to help quantify the performance benefit of the meta-tracing JIT.

The figure below shows the performance of each ISS executing applications from the SPEC CINT2006 benchmark suite [h]. Benchmarks were run to completion on the high-performance DBT-ISSs (simit-jit, pydgin-jit, and QEMU), but were terminated after only 10 billion simulated instructions for the non-JITed interpretive ISSs (these would require many hours, in some cases days, to run to completion). Simulation performance is measured in MIPS [i] and plotted on a log scale due to the wide variance in performance. The WHMEAN group summarizes each ISS's performance across all benchmarks using the weighted harmonic mean.

A few points to take away from these results:

  • ISSs without JITs (gem5, simit-nojit, and pydgin-nojit) demonstrate relatively consistent performance across applications, whereas ISSs with JITs (simit-jit, pydgin-jit, and QEMU) demonstrate much greater performance variability from application-to-application.
  • The gem5 atomic model demonstrates particularly miserable performance, only 2-3 MIPS!
  • QEMU lives up to its reputation as a gold-standard for simulator performance, leading the pack on nearly every benchmark and reaching speeds of 240-1120 MIPS.
  • pydgin-jit is able to outperform simit-jit on four of the applications, including considerable performance improvements of 1.44–1.52× for the applications 456.hmmer, 462.libquantum, and 471.omnetpp (managing to even outperform QEMU on 471.omnetpp).
  • simit-jit is able to obtain much more consistent performance (230-459 MIPS across all applications) than pydgin-jit (9.6-659 MIPS). This is due to simit-jit's page-based approach to JIT optimization compared to pydgin-jit's tracing-based approach.
  • 464.h264ref displays particularly bad pathological behavior in Pydgin’s tracing JIT and is the only application to perform worse on pydgin-jit than pydgin-nojit (9.6 MIPS vs. 21 MIPS).

The pathological behavior demonstrated by 464.h264ref was of particular concern because it caused pydgin-jit to perform even worse than having no JIT at all. RPython JIT logs indicated that the reason for this performance degradation was a large number of tracing aborts due to JIT traces growing too long. However, time limitations before the publication deadline prevented us from investigating this issue thoroughly.

Since the deadline we've applied some minor bug fixes and made some small improvements in the memory representation. More importantly, we've addressed the performance degradation in 464.h264ref by increasing trace lengths for the JIT. Below we show how the performance of 464.h264ref changes as the trace_limit parameter exposed by the RPython JIT is varied from the default size of 6000 operations.

By quadrupling the trace limit we achieve an 11x performance improvement in 464.h264ref. The larger trace limit allows the JIT to optimize long code paths that were previously triggering trace aborts, greatly helping amortize the costs of tracing. Note that arbitrarily increasing this limit can potentially hurt performance if longer traces are not able to detect optimizable code sequences.

After performing similar experiments across the applications in the SPEC CINT2006 benchmark suite, we settled on a trace limit of 400,000 operations. In the figure below we show how the updated Pydgin ISS (pydgin-400K) improves performance across all benchmarks and fixes the performance degradation previously seen in 464.h264ref. Note that the non-JITted simulators have been removed for clarity, and simulation performance is now plotted on a linear scale to more clearly distinguish the performance gap between each ISS.

With these improvements, we are now able to beat simit-jit on all but two benchmarks. In future work we hope to further close the gap with QEMU as well.

Conclusions and Future Work

Pydgin demonstrates that the impressive work put into the RPython translation toolchain, designed to simplify the process of building fast dynamic-language VMs, can also be leveraged to build fast instruction set simulators. Our prototype ARMv5 ISS shows that Pydgin can generate ISSs with performance competitive to SimIt-ARM while also providing a more productive development experience: RPython allowed us to develop Pydgin with only four person-months of work. Another significant benefit of the Pydgin approach is that any performance improvements applied to the RPython translation toolchain immediately benefit Pydgin ISSs after a simple software download and retranslation. This allows Pydgin to track the continual advances in JIT technology introduced by the PyPy development team.

Pydgin is very much a work in progress. There are many features we would like to add, including:

  • more concise syntax for accessing arbitrary instruction bits
  • support for other C libraries such as glibc, uClibc, and musl (we currently only support binaries compiled with newlib)
  • support for self-modifying code
  • features for more productive debugging of target applications
  • ISS descriptions for other ISAs such as RISC-V, ARMv8, and x86
  • automatic generation of compilers and toolchains from Pydgin descriptions

In addition, we think there are opportunities for even greater performance improvements with more advanced techniques such as:

  • automatic generation of optimized instruction decoders
  • optimizations for floating-point intensive applications
  • multiple tracing-JITs for parallel simulation of multicore SOCs
  • a parallel JIT compilation engine as proposed by Böhm et al. [3]

We hope that Pydgin can be of use to others, so if you try it out please let us know what you think. Feel free to contact us if you find any of the above development projects interesting, or simply fork the project on GitHub and hack away!

-- Derek Lockhart and Berkin Ilbeyi


We would like to sincerely thank Carl Friedrich Bolz and Maciej Fijalkowski for their feedback on the Pydgin publication and their guidance on improving the JIT performance of our simulators. We would also like to thank for the whole PyPy team for their incredible work on the PyPy and the RPython translation toolchain. Finally, thank you to our research advisor, Prof. Christopher Batten, and the sponsors of this work which include the National Science Foundation, the Defense Advanced Research Projects Agency, and Intel Corporation.


[a]Pydgin loosely stands for [Py]thon [D]SL for [G]enerating [In]struction set simulators and is pronounced the same as “pigeon”. The name is inspired by the word “pidgin” which is a grammatically simplified form of language and captures the intent of the Pydgin embedded-ADL.
[b]Popular instruction set architectures (ISAs) include MIPs, ARM, x86, and more recently RISC-V
[c](1, 2) For a good discussion of simulators vs. emulators, please see the following post on StackOverflow:
[e]Please see the Pydgin paper for a more detailed discussion of prior work.

For more examples of Pydgin ISA specifications, please see the ISPASS paper [1] or the Pydgin source code on GitHub.

Pydgin instruction definitions for a simple MIPS-inspired ISA can be found here:

Pydgin instruction definitions for a simplified ARMv5 ISA can be found here:


gem5 is a cycle-level simulation framework that contains both functional-level (atomic) and cycle-level processor models. Although primarily used for detailed, cycle-approximate processor simulation, gem5's atomic model is a popular tool for many ISS tasks.

[h]All performance measurements were taken on an unloaded server-class machine.
[i]Millions of instructions per second.


[1](1, 2, 3)

Derek Lockhart, Berkin Ilbeyi, and Christopher Batten. "Pydgin: Generating Fast Instruction Set Simulators from Simple Architecture Descriptions with Meta-Tracing JIT Compilers." IEEE Int'l Symp. on Performance Analysis of Systems and Software (ISPASS), Mar. 2015.


Derek Lockhart, Gary Zibrat, and Christopher Batten. "PyMTL: A Unified Framework for Vertically Integrated Computer Architecture Research." 47th ACM/IEEE Int'l Symp. on Microarchitecture (MICRO-47), Dec. 2014.

[3]I. Böhm, B. Franke, and N. Topham. Generalized Just-In-Time Trace Compilation Using a Parallel Task Farm in a Dynamic Binary Translator. ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI), Jun 2011.

Wednesday, February 25, 2015

Experiments in Pyrlang with RPython

Pyrlang is an Erlang BEAM bytecode interpreter written in RPython.

It implements approximately 25% of BEAM instructions. It can support integer calculations (but not bigint), closures, exception handling, some operators to atom, list and tuple, user modules, and multi-process in single core. Pyrlang is still in development.

There are some differences between BEAM and the VM of PyPy:

  • BEAM is a register-based VM, whereas the VM in PyPy is stack-based.
  • There is no traditional call-stack in BEAM. The Y register in BEAM is similar to a call-stack, but the Y register can sometimes store some variables.
  • There are no typical language-level threads and OS-level threads in BEAM; only language-level processes, whose behavior is very similar to the actor model.

Regarding bytecode dispatch loop, Pyrlang uses a while loop to fetch instructions and operands, call the function corresponding to every instruction, and jump back to the head of the while loop. Due to the differences between the RPython call-stack and BEAM’s Y register, we decided to implement and manage the Y register by hand. On the other hand, PyPy uses RPython’s call stack to implement Python’s call stack. As a result, the function for the dispatch loop in PyPy calls itself recursively. This does not happen in Pyrlang.

The Erlang compiler (erlc) usually compiles the bytecode instructions for function invocation into CALL (for normal invocation) and CALL_ONLY (for tail recursive invocation). You can use a trampoline semantic to implement it:

  • CALL instruction: The VM pushes the current instruction pointer (or called-program counter in PyPy) to the Y register, and jumps to the destination label. When encountering a RETURN instruction, the VM pops the instruction pointer from the Y register and returns to the location of the instruction pointer to continue executing the outer function.
  • CALL_ONLY instruction: The VM simply jumps to the destination label, without any modification of the Y register. As a result, the tail recursive invocation never increases the Y register.

The current implementation only inserts the JIT hint of can_enter_jit following the CALL_ONLY instruction. This means that the JIT only traces the tail-recursive invocation in Erlang code, which has a very similar semantic to the loop in imperative programming languages like Python.

We have also written a single scheduler to implement the language level process in a single core. There is a runable queue in the scheduler. On each iteration, the scheduler pops one element (which is a process object with dispatch loop) from the queue, and executes the dispatch loop of the process object. In the dispatch loop, however, there is a counter-call “reduction” inside the dispatch loop. The reduction decrements during the execution of the loop, and when the reduction becomes 0, the dispatch loop terminates. Then the scheduler pushes that element into the runable queue again, and pops the next element for the queue, and so on.

We are planning to implement a multi-process scheduler for multi-core CPUs, which will require multiple schedulers and even multiple runable queues for each core, but that will be another story. :-)


We wrote two benchmark programs of Erlang:

  • FACT: A benchmark to calculate the factorial in a tail-recursive style, but because we haven’t implemented big int, we do a remainder calculation to the argument for the next iteration, so the number never overflows.
  • REVERSE: The benchmark creates a reversed list of numbers, such as [20000, 19999, 19998, …], and applies a bubble sort to it.


The Value of Reduction

We used REVERSE to evaluate the JIT with different values of reduction:

The X axis is the value of reduction, and the Y axis is the execution time (by second).

It seems that when the value of reduction is small, the reduction influences the performance significantly, but when reduction becomes larger, it only increases the speed very slightly. In fact, we use 2000 as the default reduction value (as well as the reduction value in the official Erlang interpreter).

Surprisingly, the trace is always generated even when the reduction is very small, such as 0, which means the dispatch loop can only run for a very limited number of iterations, and the language level process executes fewer instructions than an entire loop in one switch of the scheduler). The generated trace is almost the same, regardless of different reduction values.

Actually, the RPython JIT only cares what code it meets, but does not care who executes it, thus the JIT always generates the results above. The trace even can be shared among different threads if they execute the same code.

The overhead at low reduction value may be due to the scheduler, which switches from different processes too frequently, or from the too-frequent switching between bytecode interpreter and native code, but not from JIT itself.

Here is more explanation from Armin Rigo:

“The JIT works well because you’re using a scheme where some counter is decremented (and the soft-thread interrupted when it reaches zero) only once in each app-level loop. The soft-thread switch is done by returning to some scheduler, which will resume a different soft-thread by calling it. It means the JIT can still compile each of the loops as usual, with the generated machine code containing the decrease-and-check-for-zero operation which, when true, exits the assembler."

Fair Process Switching vs. Unfair Process Switching

We are also concerned about the timing for decreasing reduction value. In our initial version of Pyrlang, we decrease reduction value at every local function invocation, module function invocation, and BIF (built-in function) invocation, since this is what the official Erlang interpreter does. However, since the JIT in RPython basically traces the target language loop (which is the tail recursive invocation in Pyrlang) it is typically better to keep the loop whole during a switch of the language level process. We modified Pyrlang, and made the reduction decrement only occur after CALL_ONLY, which is actually the loop boundary of the target language.

Of course, this strategy may cause an “unfair” execution among language level processes. For example, if one process has only a single long-sequence code, it executes until the end of the code. On the other hand, if a process has a very short loop, it may be executed by very limited steps then be switched out by the scheduler. However, in the real world, this “unfairness” is usually considered acceptable, and is used in many VM implementations including PyPy for improving the overall performance.

We compared these two versions of Pyrlang in the FACT benchmark. The reduction decrement is quite different because there are some BIF invocations inside the loop. In the old version the process can be suspended at loop boundaries or other function invocation, but in the new version, it can be suspended only at loop boundaries.

We show that the strategy is effective, removing around 7% of the overhead. We have also compared it in REVERSE, but since there are no extra invocations inside the trace, it cannot provide any performance improvement. In the real world, we believe there is usually more than one extra invocation inside a single loop, so this strategy is effective for most cases.

Comparison with Default Erlang and HiPE

We compared the performance of Pyrlang with the default Erlang interpreter and the HiPE (High Performance Erlang) complier. HiPE is an official Erlang compiler that can compile Erlang source code to native code. The speed of Erlang programs obviously improves but loses its generality instead.

Please note that Pyrlang is still in development, so in some situations it does less work than the default Erlang interpreter, such as not checking integer overflow when dealing with big integer, and not checking and adding locks when accessing message queues in the language-level process, so is therefore faster. The final version of Pyrlang may be slower.

We used the two benchmark programs above, and made sure both of them are executed for more than five seconds to cover the JIT warm-up time for RPython. The experiment environment is a OS X 10.10 machine with 3.5GHZ 6-core Intel Xeon E5 CPU and 14GB 1866 MHz DDR3 ECC memory.

Let’s look at the result of FACT. The graph shows that Pyrlang runs 177.41% faster on average than Erlang, and runs at almost the same speed as HiPE. However, since we haven’t implemented big integer in Pyrlang, the arithmetical operators do not do any extra overflow checking. It is reasonable that the final version for Pyrlang will be slower than the current version and HiPE.

As for REVERSE, the graph shows that Pyrlang runs 45.09% faster than Erlang, but 63.45% slower than HiPE on average. We think this is reasonable because there are only few arithmetical operators in this benchmark so the speeds of these three implementations are closer. However, we observed that at the scale of 40,000, the speed of Pyrlang slowed down significantly (111.35% slower than HiPE) compared with the other two scales (56.38% and 22.63% slower than HiPE).

Until now we can only hypothesize why Pyrlang slows down at that scale. We guess that the overhead might be from GC. This is because the BEAM bytecode provides some GC hints to help the default Erlang compiler to perform some GC operations immediately. For example, using GC_BIF instead of a BIF instruction tells the VM that there may be a GC opportunity, and tells the VM how many live variables should be around one instruction. In Pyrlang we do not use these kinds of hints but rely on RPython’s GC totally. When there are a huge number of objects during runtime, (as for REVERSE, it should be the Erlang list object) the speed therefore slows down.

Ruochen Huang

Monday, February 23, 2015

linalg support in pypy/numpy


PyPy's numpy support has matured enough that it can now support the lapack/blas libraries through the numpy.linalg module. To install the version of numpy this blog post refers to, install PyPy version 2.5.0 or newer, and run this:

pypy -m pip install git+

This update is a major step forward for PyPy's numpy support. Many of the basic matrix operations depend on linalg, even matplotlib requires it to display legends (a pypy-friendly version of matplotlib 1.3 is available at

A number of improvements and adaptations, some of which are in the newly-released PyPy 2.5.0, made this possible:
  • Support for an extended frompyfunc(), which in the PyPy version supports much of the ufunc API (signatures, multiple dtypes) allowing creation of pure-python, jit-friendly ufuncs. An additional keyword allows choosing between out = func(in) or func(in, out) ufunc signatures. More explanation follows.
  • Support for GenericUfuncs via PyPy's (slow) capi-compatibility layer. The underlying mechanism actually calls the internal implementation of frompyfunc().
  • A cffi version of _umath_linalg. Since cffi uses dlopen() to call into shared objects, we added support in the numpy build system to create non-python shared libraries from source code in the numpy tree. We also rewrote parts of the c-based _umath_linalg.c.src in python, renamed numpy's umath_linalg capi module to umath_linag_capi, and use it as a shared object through cffi.


We have not completely implemented all the linalg features. dtype resolution via casting is missing, especially for complex ndarrays, which leads to slight numerical errors where numpy uses a more precise type for intermediate calculations. Other missing features in PyPy's numpy support may have implications for complete linalg support.

Some OSX users have noticed they need to update pip to version 6.0.8 to overcome a regression in pip, and it is not clear if we support all combinations of blas/lapack implementations on all platforms.

Over the next few weeks we will be ironing out these issues.


A simple benchmark is shown below, but let's state the obvious: PyPy's JIT and the iterators built into PyPy's ndarray implementation will in most cases be no faster than CPython's numpy. The JIT can help where there is a mixture of python and numpy-array code. We do have plans to implement lazy evaluation and to further optimize PyPy's support for numeric python, but numpy is quite good at what it does.

HowTo for PyPy's extended frompyfunc

The magic enabling blas support is a rewrite of the _umath_linalg c-based module as a cffi-python module that creates ufuncs via frompyfunc. We extended the numpy frompyfunc to allow it to function as a replacement for the generic ufunc available in numpy only through the c-api.

We start with the basic frompyfunc, which wraps a python function into a ufunc:
def times2(in0):
    return in0 * 2
ufunc = frompyfunc(times2, 1, 1)

In cpython's numpy the dtype of the result is always object, which is not implemented (yet) in PyPy, so this example will fail. While the utility of object dtypes can be debated, in the meantime we add a non-numpy-compatible keyword argument dtypes to frompyfunc. If dtype=['match'] the output dtype will match the dtype of the first input ndarray:

ufunc = frompyfunc(times2, 1, 1, dtype=['match'])
ai = arange(24).reshape(3, 4, 2)
ao = ufunc(ai)
assert  (ao == ai * 2).all()

I hear you ask "why is the dtypes keyword argument a list?" This is so we can support the Generalized Universal Function API, which allows specifying a number of specialized functions and the input-output dtypes each specialized function accepts.
Note that the function feeds the values of ai one at a time, the function operates on scalar values. To support more complicated ufunc calls, the generalized ufunc API allows defining a signature, which specifies the layout of the ndarray inputs and outputs. So we extended frompyfunc with a signature keyword as well.
We add one further extension to frompyfunc: we allow a Boolean keyword stack_inputs to specify the argument layout of the function itself. If the function is of the form:
out0, out1, ... = func(in0, in1,...)

then stack_inputs is False. If it is True the function is of the form:
func(in0, in1, ... out0, out1, ...)

Here is a complete example of using frompyfunc to create a ufunc, based on this link:
def times2(in_array, out_array):
    in_flat = in_array.flat
    out_flat = out_array.flat
    for i in range(in_array.size):
        out_flat[i] = in_flat[i] * 2
ufunc = frompyfunc([times2, times2], 1, 1,
                dtypes=[dtype(int), dtype(int),
                        dtype(float), dtype(float),
ai = arange(10, dtype=int)
ai2 = ufunc(ai)
assert all(ai2 == ai * 2)

Using this extended syntax, we rewrote the lapack calls into the blas functions in pure python, no c needed. Benchmarking this approach actually was much slower than using the upstream umath_linalg module via cpyext, as can be seen in the following benchmarks. This is due to the need to copy c-aligned data into Fortran-aligned format. Our __getitem__ and __setitem__ iterators are not as fast as pointer arithmetic in C. So we next tried a hybrid approach: compile and use numpy's umath_linalg python module as a shared object, and call the optimized specific wrapper function from it.


Here are some benchmarks, running a tight loop of the different versions of linalg.inv(a), where a is a 10x10 double ndarray. The benchmark ran on an i7 processor running ubuntu 14.04 64 bit:
Impl. Time after warmup
CPython 2.7 + numpy + lapack 8.9 msec/1000 loops
PyPy 2.5.0 + numpy + lapack via cpyext 8.6 msec/1000 loops
PyPy 2.5.0 + numpy + lapack via pure python + cffi 19.9 msec/1000 loops
PyPy 2.5.0 + numpy + lapack via python + c + cffi 9.5 msec/1000 loops

While no general conclusions may be drawn from a single micro-benchmark, it does indicate that there is some merit in the approach taken.


PyPy's numpy now includes a working linalg module. There are still some rough corners, but hopefully we have implemented the parts you need. While the speed of the isolated linalg function is no faster than CPython and upstream numpy, it should not be significantly slower either. Your use case may see an improvement if you use a mix of python and lapack, which is the usual case.

Please let us know how it goes. We love to hear success stories too.

We still have challenges at all levels of programming,and are always looking for people willing to contribute, so stop by on IRC at #pypy.

mattip and the PyPy Team

Wednesday, February 11, 2015

NumPyPy status - January 2015

Hi Everyone

Here is what has been done in January thanks to the funding of NumPyPy, I would like to thank all the donors and tell you that you can still donate :
  • I have focused on implementing the object dtype this month, it is now possible to store objects inside ndarrays using the object dtype
  • It is also possible to add an object ndarray to any other ndarray (implementing other operators is trivial)
The next things I plan on working on next are :
  • Implementing the missing operations for object arrays
  • Implementing garbage collection support for object arrays (currently, storing an object inside an ndarray doesn't keep the object alive)
  • Packaging NumPyPy on PyPI

Tuesday, February 3, 2015

PyPy 2.5.0 released

PyPy 2.5.0 - Pincushion Protea

We’re pleased to announce PyPy 2.5, which contains significant performance enhancements and bug fixes.
You can download the PyPy 2.5.0 release here:
We would like to thank our donors for the continued support of the PyPy project, and for those who donate to our three sub-projects, as well as our volunteers and contributors (10 new commiters joined PyPy since the last release). We’ve shown quite a bit of progress, but we’re slowly running out of funds. Please consider donating more, or even better convince your employer to donate, so we can finish those projects! The three sub-projects are:
  • Py3k (supporting Python 3.x): We have released a Python 3.2.5 compatible version
    we call PyPy3 2.4.0, and are working toward a Python 3.3 compatible version
  • STM (software transactional memory): We have released a first working version, and continue to try out new promising paths of achieving a fast multithreaded Python
  • NumPy which requires installation of our fork of upstream numpy, available on bitbucket

What is PyPy?

PyPy is a very compliant Python interpreter, almost a drop-in replacement for CPython 2.7. It’s fast (pypy and cpython 2.7.x performance comparison) due to its integrated tracing JIT compiler.
This release supports x86 machines on most common operating systems (Linux 32/64, Mac OS X 64, Windows, and OpenBSD), as well as newer ARM hardware (ARMv6 or ARMv7, with VFPv3) running Linux.
While we support 32 bit python on Windows, work on the native Windows 64 bit python is still stalling, we would welcome a volunteer to handle that.


  • The past months have seen pypy mature and grow, as rpython becomes the goto solution for writing fast dynamic language interpreters. Our separation of rpython and the python interpreter PyPy is now much clearer in the PyPy documentation and we now have separate RPython documentation.
  • We have improved warmup time as well as jitted code performance: more than 10% compared to pypy-2.4.0. We no longer zero-out memory allocated in the gc nursery by default, work that was started during a GSoC.
  • Passing objects between C and PyPy has been improved. We are now able to pass raw pointers to C (without copying) using pinning. This improves I/O; benchmarks that use networking intensively improved by about 50%. File() operations still need some refactoring but are already showing a 20% improvement on our benchmarks. Let us know if you see similar improvements.
  • Our integrated numpy support gained much of the GenericUfunc api in order to support the lapack/blas linalg module of numpy. This dovetails with work in the pypy/numpy repository to support linalg both through the (slower) cpyext capi interface and also via (the faster) pure python cffi interface, using an extended frompyfunc() api. We will soon post a seperate blog post specifically about linalg and PyPy.
  • Dictionaries are now ordered by default, see the blog post
  • Our nightly translations use –shared by default, including on OS/X and linux
  • We now more carefully handle errno (and GetLastError, WSAGetLastError) tying the handlers as close as possible to the external function call, in non-jitted as well as jitted code.
  • Issues reported with our previous release were resolved after reports from users on our issue tracker at or on IRC at #pypy.
We have further improvements on the way: rpython file handling, finishing numpy linalg compatibility, numpy object dtypes, a better profiler, as well as support for Python stdlib 2.7.9.
Please try it out and let us know what you think. We especially welcome success stories, we know you are using PyPy, please tell us about it!
The PyPy Team

Thursday, January 22, 2015

Faster, more memory efficient and more ordered dictionaries on PyPy

Hello everyone!

As of today, we merged the latest branch that brings better dictionaries to PyPy by default. The work is based on an idea by Raymond Hettinger on python-dev, with prior work done notably in Java.  It was done by Maciej Fijałkowski and Armin Rigo, with Laurence Tratt recently prodding us to finish it.  (Earlier work going in a similar direction include Alex Gaynor's work on ordered dicts in Topaz, which was also used in the Hippy VM.  Each of these pieces of work is itself based on the original dict implementation in RPython, whose origins fade in the Subversion prehistory of PyPy.)  Coincidentally, a very similar idea has been implemented in Zend PHP very recently. Zend implementation description.

This post covers the basics of design and implementation as well as some basic benchmarks.

Dictionaries are now ordered!

One surprising part is that the new design, besides being more memory efficient, is ordered by design: it preserves the insertion order.  This is not forbidden by the Python language, which allows any order.  It makes the collections.OrderedDict subclass much faster than before: it is now a thin subclass of dict.  Obviously, we recommend that any portable Python program continues to use OrderedDict when ordering is important.  Note that a non-portable program might rely on more: for example, a **keywords argument now receives the keywords in the same order as the one in which they were given in the call.  (Whether such a thing might be called a language design change or not is a bit borderline.)  The point is that Python programs that work on CPython or previous versions of PyPy should continue to work on PyPy.

There is one exception, though.  The iterators of the OrderedDict subclass are now working just like the ones of the dict builtin: they will raise RuntimeError when iterating if the dictionary was modified.  In the CPython design, the class OrderedDict explicitly doesn't worry about that, and instead you get some result that might range from correct to incorrect to crashes (i.e. random Python exceptions).

Original PyPy dictionary design

Originally, PyPy dictionaries, as well as CPython dictionaries are implemented as follows (simplified view):

struct dict {
   long num_items;
   dict_entry* items;   /* pointer to array */

struct dict_entry {
   long hash;
   PyObject* key;
   PyObject* value;

Where items is a sparse array, with 1/3 to 1/2 of the items being NULL. The average space occupied by a dictionary is 3 * WORD * 12/7 plus some small constant (the smallest dict has 8 entries, which is 8 * 3 * WORD + 2 * WORD = 26 WORDs).

New PyPy dictionary design

The new PyPy dictionary is split in two arrays:

struct dict {
    long num_items;
    variable_int *sparse_array;
    dict_entry* compact_array;

struct dict_entry {
    long hash;
    PyObject *key;
    PyObject *value;

Here, compact_array stores all the items in order of insertion, while sparse_array is a 1/2 to 2/3 full array of integers. The integers themselves are of the smallest size necessary for indexing the compact_array. So if compact_array has less than 256 items, then sparse_array will be made of bytes; if less than 2^16, it'll be two-byte integers; and so on.

This design saves quite a bit of memory. For example, on 64bit systems we can, but almost never, use indexing of more than 4 billion elements; and for small dicts, the extra sparse_array takes very little space.  For example a 100 element dict, would be on average for the original design on 64bit: 100 * 12/7 * WORD * 3 =~ 4100 bytes, while on new design it's 100 * 12/7 + 3 * WORD * 100 =~ 2600 bytes, quite a significant saving.

GC friendliness

The obvious benefit of having more compact dictionaries is an increased cache friendliness. In modern CPUs cache misses are much more costly than doing additional simple work, like having an additional level of (in-cache) indirection. Additionally, there is a GC benefit coming from it. When doing a minor collection, the GC has to visit all the GC fields in old objects that can point to young objects. In the case of large arrays, this can prove problematic since the array grows and with each minor collection we need to visit more and more GC pointers. In order to avoid it, large arrays in PyPy employ a technique called "card marking" where the GC only visits "cards" or subsets of arrays that were modified between collections. The problem with dictionaries was that by design modifications in a dictionary occur randomly, hence a lot of cards used to get invalidated. In the new design, however, new items are typically appended to the compact_array, hence invalidate much fewer cards --- which improves GC performance.  (The new sparse_array is an array of integers, so it does not suffer from the same problems.)


Deleting entries from dictionaries is not very common, but important in a few use cases.  To preserve order, when we delete an entry, we mark the entry as removed but don't otherwise shuffle the remaining entries.  If we repeat this operation often enough, there will be a lot of removed entries in the (originally compact) array.  At this point, we need to do a "packing" operation, which moves all live entries to the start of the array (and then reindexes the sparse array, as the positions changed).  This works well, but there are use cases where previously no reindexing was ever needed, so it makes these cases a bit slower (for example when repeatedly adding and removing keys in equal number).


The PyPy speed benchmarks show mostly small effect, see changes. The microbenchmarks that we did show large improvements on large and very large dictionaries (particularly, building dictionaries of at least a couple 100s of items is now twice faster) and break-even on small ones (between 20% slower and 20% faster depending very much on the usage patterns and sizes of dictionaries). The new dictionaries enable various optimization possibilities which we're going to explore in the near future.

fijal, arigo and the PyPy team

Thursday, January 15, 2015

Leysin Winter Sprint (20-28th February 2015)

The next PyPy sprint will be in Leysin, Switzerland, for the tenth time. This is a fully public sprint: newcomers and topics other than those proposed below are welcome.

Goals and topics of the sprint

The details depend on who is here and ready to work. We might touch topics such as:

  • cleaning up the optimization step in the JIT, change the register allocation done by the JIT's backend, or improvements to the warm-up time

  • STM (Software Transaction Memory), notably: try to come up with benchmarks, and measure them carefully in order to test and improve the conflict reporting tools, and more generally to figure out how practical it is in large projects to avoid conflicts

  • vmprof - a statistical profiler for CPython and PyPy work, including making it more user friendly.

  • Py3k (Python 3.x support), NumPyPy (the numpy module)

  • added: cffi 1.0, trying out pygame+cffi on Raspberry Pi devices

  • And as usual, the main side goal is to have fun in winter sports :-) We can take a day off for ski.

Exact times

For a change, and as an attempt to simplify things, I specified the dates as 20-28 Februrary 2015, where 20 and 28 are travel days. We will work full days between the 21 and the 27. You are of course allowed to show up for a part of that time only, too.

Location and Accomodation

Leysin, Switzerland, "same place as before". Let me refresh your memory: both the sprint venue and the lodging will be in a very spacious pair of chalets built specifically for bed & breakfast: Ermina. The place has a good ADSL Internet connection with wireless installed. You can of course arrange your own lodging anywhere (as long as you are in Leysin, you cannot be more than a 15 minutes walk away from the sprint venue), but I definitely recommend lodging there too -- you won't find a better view anywhere else (though you probably won't get much worse ones easily, either :-)

Please confirm that you are coming so that we can adjust the reservations as appropriate. In the past, the rates were around 60 CHF a night all included in 2-person rooms, with breakfast. Now, the rooms available are either single-person (or couple), or rooms for 3 persons. The latter choice is recommended and should be under 60 CHF per person.

Please register by Mercurial, or on the pypy-dev mailing list if you do not yet have check-in rights.

You need a Swiss-to-(insert country here) power adapter. There will be some Swiss-to-EU adapters around, and at least one EU-format power strip.

Friday, November 28, 2014

September donations and thank you to the Python Software Foundation!

Hello everyone!

We would like to show you a short update on the PyPy funding. We gathered a total of $15,986 in the month of September and as per earlier agreement, the Python Software Foundation donated $10,000 to PyPy. We would like to thank everyone participating and the PSF in particular for supporting the PyPy project and making our work possible!

We've been working hard on the goals outlined in the funding proposals.

  • PyPy Python 3 support has been in beta for a while and it's already being used by many people, as seen per the number of reported bugs. We're currently supporting 3.2, planning on moving towards 3.4 in the future.
  • Software Transactional Memory has been a successful research project, with first real world results shown during the Warsaw sprint.
  • More detailed update on numpy will be published soon. A little spoiler is that we're planning on addressing matplotlib, scipy and the larger ecosystem to some extent. Stay tuned!

Again, thanks to everyone who donated and happy Thanksgiving to everyone on that side of the world!

fijal and the entire PyPy team

Monday, November 17, 2014

Tornado without a GIL on PyPy STM

This post is by Konstantin Lopuhin, who tried PyPy STM during the Warsaw sprint.

Python has a GIL, right? Not quite - PyPy STM is a python implementation without a GIL, so it can scale CPU-bound work to several cores. PyPy STM is developed by Armin Rigo and Remi Meier, and supported by community donations. You can read more about it in the docs.

Although PyPy STM is still a work in progress, in many cases it can already run CPU-bound code faster than regular PyPy, when using multiple cores. Here we will see how to slightly modify Tornado IO loop to use transaction module. This module is described in the docs and is really simple to use - please see an example there. An event loop of Tornado, or any other asynchronous web server, looks like this (with some simplifications):

while True:
    for callback in list(self._callbacks):
    event_pairs = self._impl.poll()
    while self._events:
        fd, events = self._events.popitem()
        handler = self._handlers[fd]
        self._handle_event(fd, handler, events)

We get IO events, and run handlers for all of them, these handlers can also register new callbacks, which we run too. When using such a framework, it is very nice to have a guaranty that all handlers are run serially, so you do not have to put any locks. This is an ideal case for the transaction module - it gives us guaranties that things appear to be run serially, so in user code we do not need any locks. We just need to change the code above to something like:

while True:
    for callback in list(self._callbacks):
        transaction.add(                # added
            self._run_callback, callback)                   # added
    event_pairs = self._impl.poll()
    while self._events:
        fd, events = self._events.popitem()
        handler = self._handlers[fd]
        transaction.add(                # added
            self._handle_event, fd, handler, events)                   # added

The actual commit is here, - we had to extract a little function to run the callback.

Part 1: a simple benchmark: primes

Now we need a simple benchmark, lets start with this - just calculate a list of primes up to the given number, and return it as JSON:

def is_prime(n):
    for i in xrange(2, n):
        if n % i == 0:
            return False
    return True

class MainHandler(tornado.web.RequestHandler):
    def get(self, num):
        num = int(num)
        primes = [n for n in xrange(2, num + 1) if is_prime(n)]
        self.write({'primes': primes})

We can benchmark it with siege:

siege -c 50 -t 20s http://localhost:8888/10000

But this does not scale. The CPU load is at 101-104 %, and we handle 30 % less request per second. The reason for the slowdown is STM overhead, which needs to keep track of all writes and reads in order to detect conflicts. And the reason for using only one core is, obviously, conflicts! Fortunately, we can see what this conflicts are, if we run code like this (here 4 is the number of cores to use):

PYPYSTM=stm.log ./ 4

Then we can use to analyse this log. It lists the most expensive conflicts:

14.793s lost in aborts, 0.000s paused (1258x STM_CONTENTION_INEVITABLE)
File "/home/ubuntu/tornado-stm/tornado/tornado/", line 455, in __init__
    self._start_time = time.time()
File "/home/ubuntu/tornado-stm/tornado/tornado/", line 455, in __init__
    self._start_time = time.time()

There are only three kinds of conflicts, they are described in stm source, Here we see that two threads call into external function to get current time, and we can not rollback any of them, so one of them must wait till the other transaction finishes. For now we can hack around this by disabling this timing - this is only needed for internal profiling in tornado.

If we do it, we get the following results (but see caveats below):

Impl. req/s
PyPy 2.4 14.4
CPython 2.7 3.2
PyPy-STM 1 9.3
PyPy-STM 2 16.4
PyPy-STM 3 20.4
PyPy STM 4 24.2

As we can see, in this benchmark PyPy STM using just two cores can beat regular PyPy! This is not linear scaling, there are still conflicts left, and this is a very simple example but still, it works!

But its not that simple yet :)

First, these are best-case numbers after long (much longer than for regular PyPy) warmup. Second, it can sometimes crash (although removing old pyc files fixes it). Third, benchmark meta-parameters are also tuned.

Here we get relatively good results only when there are a lot of concurrent clients - as a results, a lot of requests pile up, the server is not keeping with the load, and transaction module is busy with work running this piled up requests. If we decrease the number of concurrent clients, results get slightly worse. Another thing we can tune is how heavy is each request - again, if we ask primes up to a lower number, then less time is spent doing calculations, more time is spent in tornado, and results get much worse.

Besides the time.time() conflict described above, there are a lot of others. The bulk of time is lost in these two conflicts:

14.153s lost in aborts, 0.000s paused (270x STM_CONTENTION_INEVITABLE)
File "/home/ubuntu/tornado-stm/tornado/tornado/", line 1082, in compute_etag
    hasher = hashlib.sha1()
File "/home/ubuntu/tornado-stm/tornado/tornado/", line 1082, in compute_etag
    hasher = hashlib.sha1()

13.484s lost in aborts, 0.000s paused (130x STM_CONTENTION_WRITE_READ)
File "/home/ubuntu/pypy/lib_pypy/", line 164, in _run_thread

The first one is presumably calling into some C function from stdlib, and we get the same conflict as for time.time() above, but is can be fixed on PyPy side, as we can be sure that computing sha1 is pure.

It is easy to hack around this one too, just removing etag support, but if we do it, performance is much worse, only slightly faster than regular PyPy, with the top conflict being:

83.066s lost in aborts, 0.000s paused (459x STM_CONTENTION_WRITE_WRITE)
File "/home/arigo/hg/pypy/stmgc-c7/lib-python/2.7/", line 70, in __contains__
File "/home/arigo/hg/pypy/stmgc-c7/lib-python/2.7/", line 70, in __contains__

Comment by Armin: It is unclear why this happens so far. We'll investigate...

The second conflict (without etag tweaks) originates in the transaction module, from this piece of code:

while True:
    counter[0] += 1

Comment by Armin: This is a conflict in the transaction module itself; ideally, it shouldn't have any, but in order to do that we might need a little bit of support from RPython or C code. So this is pending improvement.

Tornado modification used in this blog post is based on 3.2.dev2. As of now, the latest version is 4.0.2, and if we apply the same changes to this version, then we no longer get any scaling on this benchmark, and there are no conflicts that take any substantial time.

Comment by Armin: There are two possible reactions to a conflict. We can either abort one of the two threads, or (depending on the circumstances) just pause the current thread until the other one commits, after which the thread will likely be able to continue. The tool ```` did not report conflicts that cause pauses. It has been fixed very recently. Chances are that on this test it would report long pauses and point to locations that cause them.

Part 2: a more interesting benchmark: A-star

Although we have seen that PyPy STM is not all moonlight and roses, it is interesting to see how it works on a more realistic application. is a simple game where several players move on a map (represented as a list of lists of integers), build and destroy walls, and ask server to give them shortest paths between two points using A-star search, adopted from ActiveState recipie.

The benchmark is simulating players, and tries to put the main load on A-star search, but also does some wall building and destruction. There are no locks around map modifications, as normal tornado is executing all callbacks serially, and we can keep this guaranty with atomic blocks of PyPy STM. This is also an example of a program that is not trivial to scale to multiple cores with separate processes (assuming more interesting shared state and logic).

This benchmark is very noisy due to randomness of client interactions (also it could be not linear), so just lower and upper bounds for number of requests are reported

Impl. req/s
PyPy 2.4 5 .. 7
CPython 2.7 0.5 .. 0.9
PyPy-STM 1 2 .. 4
PyPy STM 4 2 .. 6

Clearly this is a very bad benchmark, but still we can see that scaling is worse and STM overhead is sometimes higher. The bulk of conflicts come from the transaction module (we have seen it above):

91.655s lost in aborts, 0.000s paused (249x STM_CONTENTION_WRITE_READ)
File "/home/ubuntu/pypy/lib_pypy/", line 164, in _run_thread

Although it is definitely not ready for production use, you can already try to run things, report bugs, and see what is missing in user-facing tools and libraries.

Benchmarks setup:

Wednesday, November 5, 2014

PyPy IO improvements

Hello everyone!

We've wrapped up the Warsaw sprint, so I would like to describe some branches which have been recently merged and which improved the I/O and the GC: gc_no_cleanup_nursery and gc-incminimark-pinning.

The first branch was started by Wenzhu Man for her Google Summer of Code and finished by Maciej Fijałkowski and Armin Rigo. The PyPy GC works by allocating new objects in the young object area (the nursery), simply by incrementing a pointer. After each minor collection, the nursery has to be cleaned up. For simplicity, the GC used to do it by zeroing the whole nursery.

This approach has bad effects on the cache, since you zero a large piece of memory at once and do unnecessary work for things that don't require zeroing like large strings. We mitigated the first problem somewhat with incremental nursery zeroing, but this branch removes the zeroing completely, thus improving the string handling and recursive code (since jitframes don't requires zeroed memory either). I measured the effect on two examples: a recursive implementation of fibonacci and gcbench, to measure GC performance.

The results for fibonacci and gcbench are below (normalized to cpython 2.7). Benchmarks were run 50 times each (note that the big standard deviation comes mostly from the warmup at the beginning, true figures are smaller):

benchmark CPython PyPy 2.4 PyPy non-zero
fibonacci 4.8+-0.15 (1.0x) 0.59+-0.07 (8.1x) 0.45+-0.07 (10.6x)
gcbench 22+-0.36 (1.0x) 1.34+-0.28 (16.4x) 1.02+-0.15 (21.6x)

The second branch was done by Gregor Wegberg for his master thesis and finished by Maciej Fijałkowski and Armin Rigo. Because of the way it works, the PyPy GC from time to time moves the objects in memory, meaning that their address can change. Therefore, if you want to pass pointers to some external C function (for example, write(2) or read(2)), you need to ensure that the objects they are pointing to will not be moved by the GC (e.g. when running a different thread). PyPy up to 2.4 solves the problem by copying the data into or from a non-movable buffer, which is obviously inefficient. The branch introduce the concept of "pinning", which allows us to inform the GC that it is not allowed to move a certain object for a short period of time. This introduces a bit of extra complexity in the garbage collector, but improves the I/O performance quite drastically, because we no longer need the extra copy to and from the non-movable buffers.

In this benchmark, which does I/O in a loop, we either write a number of bytes from a freshly allocated string into /dev/null or read a number of bytes from /dev/full. I'm showing the results for PyPy 2.4, PyPy with non-zero-nursery and PyPy with non-zero-nursery and object pinning. Those are wall times for cases using and, normalized against CPython 2.7.

Benchmarks were done using PyPy 2.4 and revisions 85646d1d07fb for non-zero-nursery and 3d8fe96dc4d9 for non-zero-nursery and pinning. The benchmarks were run once, since the standard deviation was small.

The Y axis is speed, normalized to CPython, the more the better

What we can see is that and os.write both improved greatly and outperforms CPython now for each combination. file operations are a little more tricky, and while those branches improved the situation a bit, the improvement is not as drastic as in os versions. It really should not be the case and it showcases how our file buffering is inferior to CPython. We plan on removing our own buffering and using FILE* in C in the near future, so we should outperform CPython on those too (since our allocations are cheaper). If you look carefully in the benchmark, the write function is copied three times. This hack is intended to avoid JIT overspecializing the assembler code, which happens because the buffering code was written way before the JIT was done. In fact, our buffering is hilariously bad, but if stars align correctly it can be JIT-compiled to something that's not half bad. Try removing the hack and seeing how the performance of the last benchmark drops :-) Again, this hack should be absolutely unnecessary once we remove our own buffering, stay tuned for more.