<!DOCTYPE html> <html> <head> <title>PyPy - 2nd Call for donations - Transactional Memory in PyPy</title> <meta http-equiv="content-language" content="en" /> <meta http-equiv="content-type" content="text/html; charset=utf-8" /> <meta name="author" content="PyPy Team" /> <meta name="description" content="PyPy" /> <meta name="copyright" content="MIT" /> <meta name="document-rating" content="general" /> <link rel="stylesheet" type="text/css" media="screen" title="default" href="css/site.css" /> <link rel="alternate" type="application/rss+xml" title="RSS Feed for PyPy" href="http://feeds.feedburner.com/PyPyStatusBlog" /> <link rel="stylesheet" type="text/css" href="css/jquery-ui-1.8.14.custom.css" /> <script type="text/javascript" src="https://use.typekit.com/hdt8sni.js"></script> <script type="text/javascript">try{Typekit.load();}catch(e){}</script> <script type="text/javascript" src="https://ajax.googleapis.com/ajax/libs/jquery/1.4.2/jquery.min.js"></script> <script type="text/javascript" src="js/jquery-ui-1.8.14.custom.min.js"></script> <script type="text/javascript" src="js/script2.js?bust=1"></script> </head> <body> <script type="text/javascript"> var _gaq = [['_setAccount', 'UA-7778406-3'], ['_trackPageview']]; if (document.location.protocol !== 'file:') { (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; (document.getElementsByTagName('head')[0] || document.getElementsByTagName('body')[0]).appendChild(ga); })(); } </script> <div id="body-outer"><div id="body-inner"><div id="body" class="clearfix"> <div id="header"> <div id="menu-follow"> <div><a href="https://bitbucket.org/pypy/pypy"><img src="https://www.selenic.com/hg-logo/logo-droplets-25.png" width="14px" height="14px" /></a></div> <div><a href="http://feeds.feedburner.com/PyPyStatusBlog" title="Subscribe to the RSS Feed"><img src="https://feedburner.google.com/fb/lib/images/icons/feed-icon-12x12-orange.gif" alt="Subscribe to the RSS Feed" width="14px" height="14px" /></a></div> </div> <div id="logo"><a href="http://pypy.org"><img src="image/pypy-logo.png" alt="PyPy" height="110px" /></a></div> <hr class="clear-left" /> <div id="menu-sub"> <a href="index.html">Home</a> <span class="menu-sub-sep"> | </span> <a href="features.html">What is PyPy?</a> <span class="menu-sub-sep"> | </span> <a href="download.html">Download</a> <span class="menu-sub-sep"> | </span> <a href="compat.html">Compatibility</a> <span class="menu-sub-sep"> | </span> <a href="performance.html">Performance</a> <span class="menu-sub-sep"> | </span> <a href="http://doc.pypy.org">Dev Documentation</a> <span class="menu-sub-sep"> | </span> <a href="http://morepypy.blogspot.com">Blog</a> <span class="menu-sub-sep"> | </span> <a href="people.html">People</a> <span class="menu-sub-sep"> | </span> <a href="contact.html">Contact</a> <br /> </div> <hr class="clear" /> </div> <div id="content"> <div> <div id="main"> <h1 class="title">2nd Call for donations - Transactional Memory in PyPy</h1> <p>UPDATE (September 2016):</p> <blockquote> <p><strong>This call for donation is closed. Thank you everybody for contributing!</strong></p> <p><em>We have actually more money in the pot that we can use in the near future. As it turns out, STM is a hard researchy topic. Remi Meier is still actively working on this topic as part of his Ph.D. thesis. Armin Rigo thinks more fundamental work is needed, which will eventually lead us to a next generation attempt.</em></p> </blockquote> <hr class="docutils" /> <br><br><ul class="download-menu simple"> <li><a class="reference internal" href="#preamble">Preamble</a></li> <li><a class="reference internal" href="#introduction">Introduction</a></li> <li><a class="reference internal" href="#in-more-detail">In more detail</a><ul> <li><a class="reference internal" href="#what-is-the-global-interpreter-lock">What is the Global Interpreter Lock?</a></li> <li><a class="reference internal" href="#what-is-transactional-memory">What is Transactional Memory?</a></li> <li><a class="reference internal" href="#hardware-vs-software-transactional-memory">Hardware vs Software Transactional Memory</a></li> <li><a class="reference internal" href="#why-do-tm-with-pypy-instead-of-cpython">Why do TM with PyPy instead of CPython?</a></li> <li><a class="reference internal" href="#alternatives">Alternatives</a></li> <li><a class="reference internal" href="#platforms-other-than-the-x86-64-linux">Platforms other than the x86-64 Linux</a></li> <li><a class="reference internal" href="#more-readings">More readings</a></li> </ul> </li> <li><a class="reference internal" href="#work-plan-and-funding-details">Work plan and funding details</a><ul> <li><a class="reference internal" href="#goal-1">Goal 1</a></li> <li><a class="reference internal" href="#goal-2">Goal 2</a></li> <li><a class="reference internal" href="#goal-3">Goal 3</a></li> <li><a class="reference internal" href="#funding">Funding</a></li> </ul> </li> <li><a class="reference internal" href="#benefits-of-this-work-to-the-python-community-and-the-general-public">Benefits of This Work to the Python Community and the General Public</a></li> </ul> <div class="section" id="preamble"> <h1>Preamble</h1> <p>This is the second call for donations on the topic of Transactional Memory (<em>TM</em>) in PyPy, a way to run CPU-hungry Python programs in multithreaded mode. It is a follow-up on our <a class="reference external" href="http://pypy.org/tmdonate.html">first call for donations</a> from two years ago. At that time, we suggested a single-threaded slow-down of somewhere between 2x and 5x. The aim that seems now within reach is rather closer to 1.25x, i.e. running only 25% slower than the regular PyPy.</p> <p>We achieved – or overachieved – most goals laid out in the first call by a large margin, while at the same time raising only about half the money. The result of this first step is <a class="reference external" href="https://pypy.readthedocs.org/en/latest/stm.html">described in the documentation of PyPy</a>. It is a PyPy without the GIL. In the best (artificial) examples, it runs only 30% slower than a regular PyPy with the JIT.</p> <p>The present proposal is about development of the second half: first, fixing the various missing low-level optimizations (aiming for this 25%-30% figure, but for most cases rather than only special examples). Then it will most importantly focus on developing the Python-facing interface. This includes both internal things (e.g. do dictionaries need to be more TM-friendly in general?) as well as directly visible things (e.g. some profiler-like interface to explore common conflicts in a program). Finally, the third part is exploring and tweaking some existing libraries to improve their TM-friendliness (e.g. Twisted and Stackless).</p> <p>See also the <a class="reference internal" href="#update-on-htm">update on HTM</a> below.</p> </div> <div class="section" id="introduction"> <h1>Introduction</h1> <p>In the presence of today's machines with multiple processors, Python progress is lagging behind: on any CPU-constrained program, developers have a difficult choice to make. They can use in-process solutions that do not offer multi-CPU usage. In this respect, the natural choice nowadays is to use Twisted or other event-based paradigms, or systems that hide events in the control flow, like Stackless; or alternatively, they can use the existing <tt class="docutils literal">threading</tt> module, with its associated GIL and the complexities of real multi-threaded programming (locks, deadlocks, races, etc.), which make this solution less attractive. The most attractive alternative for most developers is to rely on one of various multi-process solutions that are outside the scope of the core Python language. All of them require a major restructuring of the program and often need extreme care and extra knowledge to use them.</p> <p>We propose an implemention of Transactional Memory in PyPy. This is a technique that recently came to the forefront of the multi-core scene. It promises to offer multi-core CPU usage in a single process. In particular, by modifying the core of the event systems mentioned above, we will enable the use of multiple cores, without the user needing to use explicitly the <tt class="docutils literal">threading</tt> module.</p> <p>The first proposal was launched near the start of 2012 and has covered much of the fundamental research, up to the point of getting a first version of PyPy working in a very roughly reasonable state (after collecting about USD$27'000, which is little more than half of the money that was sought; hence the present second call for donations).</p> <p>We now propose fixing the remaining issues to obtaining a really good GIL-free PyPy (described in <a class="reference internal" href="#goal-1">goal 1</a> below). We will then focus on the various new features needed to actually use multiple cores without explicitly using multithreading (<a class="reference internal" href="#goal-2">goal 2</a> below), up to and including adapting some existing framework libraries, for example Twisted, Tornado, Stackless, or gevent (<a class="reference internal" href="#goal-3">goal 3</a> below).</p> </div> <div class="section" id="in-more-detail"> <h1>In more detail</h1> <p>This is a call for financial help in implementing a version of PyPy able to use multiple processors in a single process, called PyPy-TM; and developing the APIs and libraries needed as well as enhancing commonly available frameworks to use the new feature. The developers will be Armin Rigo and Remi Meier and possibly others.</p> <p>We currently estimate the final performance goal to be a slow-down of 25% to 40% from the current non-TM PyPy; i.e. running a fully serial application would take between 1.25 and 1.40x the time it takes in a regular PyPy. This goal has been reached already in some cases, but we need to make this result more broadly applicable. We feel confident that we can reach this goal more generally: the performance of PyPy-TM running any suitable application should scale linearly or close-to-linearly with the number of processors. This means that starting with two cores, such applications should perform better than a non-TM PyPy. (All numbers presented here are comparing different versions of PyPy which all have the JIT enabled. A “suitable application” is one without many conflicts; see <a class="reference internal" href="#goal-2">goal 2</a>.)</p> <p>You will find below a sketch of the <a class="reference internal" href="#work-plan">work plan</a>. We start with a Q&A.</p> <div class="section" id="what-is-the-global-interpreter-lock"> <h2>What is the Global Interpreter Lock?</h2> <p>The GIL, or Global Interpreter Lock, is a single lock in both CPython and the regular PyPy. Every thread must acquire it in order to execute Python bytecodes. This means that both with CPython and with the regular PyPy, Python programs do not gain any benefit in term of multicore performance even if they are using threads.</p> </div> <div class="section" id="what-is-transactional-memory"> <h2>What is Transactional Memory?</h2> <p><a class="reference external" href="http://en.wikipedia.org/wiki/Transactional_memory">Transactional Memory</a> (<em>TM</em>) is a technique imported from databases: every time we want to do a change to the processors' main memory, we do it in a “transaction”. Multiple transactions can be executed in parallel by multiple cores. When a transaction is complete, we try to commit it. This might either succeed, or (if another transaction committed incompatible changes) fail. If it fails, which is hopefully rare, we need to restart the transaction from scratch.</p> <p>Transactional Memory research has progressed a lot since two years ago, notably with the introduction of Intel's <a class="reference external" href="http://en.wikipedia.org/wiki/Haswell_%28microarchitecture%29">Haswell</a> processors, which offer Hardware Transactional Memory (HTM). We discuss below why we think HTM is, so far, still not suitable for our goals.</p> </div> <div class="section" id="hardware-vs-software-transactional-memory"> <span id="update-on-htm"></span><h2>Hardware vs Software Transactional Memory</h2> <p>The idea of Transactional Memory was recently made popular by Intel's <a class="reference external" href="http://en.wikipedia.org/wiki/Haswell_%28microarchitecture%29">Haswell</a> processor (released in 2013). We could replace most of the Software Transactional Memory (STM) library currently used inside PyPy with a much smaller Hardware Transactional Memory (HTM) library based on hardware features and running on Haswell-generation processors. This has been attempted by Remi Meier recently. However, it seems that it fails to scale as we would expect it to: the current generation of HTM processors is limited to run small-scale transactions. Even the default transaction size used in PyPy-STM is often too much for HTM; and reducing this size increases overhead without completely solving the problem. Based on this experience, it seems safe to say that right now HTM-enabled processors lack the support that we need.</p> <p>Future processors might improve on various aspects. We are particularly interested in <a class="reference external" href="http://pages.cs.wisc.edu/~isca2005/papers/08A-02.PDF">Virtualizing Transactional Memory</a>, a 2005 paper that describes the limits that we're running into and how to solve them more generally. A CPU with support for the virtual memory described in this paper would certainly be better for running PyPy-HTM.</p> <p>Another issue in HTM is sub-cache-line false conflicts (conflicts caused by two independent objects that happens to live in the same cache line, which is usually 64 bytes). This is in contrast with the current PyPy-STM, which doesn't have false conflicts of this kind at all and might thus be ultimately better for very-long-running transactions. We are not aware of published research discussing issues of sub-cache-line false conflicts.</p> <p>Note that right now PyPy-STM has false conflicts within the same object, e.g. within a list or a dictionary; but we can easily do something about it (see <cite>goal 2_</cite>). Also, it might be possible in PyPy-HTM to arrange objects in memory ahead of time so that such conflicts are very rare; but we will never get a rate of exactly 0%, which might be required for very-long-running transactions.</p> </div> <div class="section" id="why-do-tm-with-pypy-instead-of-cpython"> <h2>Why do TM with PyPy instead of CPython?</h2> <p>While there have been early experiments on Hardware Transactional Memory with CPython (<a class="reference external" href="http://sabi.net/nriley/pubs/dls6-riley.pdf">Riley and Zilles (2006)</a>, <a class="reference external" href="http://www.cs.auckland.ac.nz/~fuad/parpycan.pdf">Tabba (2010)</a>), there has been none in the past few years. To the best of our knowledge, the closest is an attempt using <a class="reference external" href="http://researcher.watson.ibm.com/researcher/files/jp-ODAIRA/PPoPP2014_RubyGILHTM.pdf">Haswell on the Ruby interpreter</a>. None of these attempts tries to do the same using Software Transactional Memory. We would nowadays consider it possible to adapt our stmgc-c7 library for CPython, but it would be a lot of work, starting from changing the reference-counting garbage collection scheme. PyPy is better designed to be open to this kind of research.</p> <p>However, the best argument from an objective point of view is probably that PyPy has already implemented a Just-in-Time compiler. It is thus starting from a better position in terms of performance, particularly for the long-running kind of programs that we target here.</p> </div> <div class="section" id="alternatives"> <h2>Alternatives</h2> <p>PyPy-TM will be slower than judicious usage of existing alternatives, based on multiple processes that communicate with each other in one way or another. The counter-argument is that TM is not only a cleaner solution: there are cases in which it is not really possible to organize (or retrofit) an existing program into the particular format needed for the alternatives. In particular, small quickly-written programs don't need the additional baggage of cross-process communication; and large programs can sometimes be almost impossible to turn into multi-process versions. By contrast, we believe that TM can fit naturally into most programs, because it only requires local changes to some dispatcher; the rest of the program should work without changes.</p> </div> <div class="section" id="platforms-other-than-the-x86-64-linux"> <h2>Platforms other than the x86-64 Linux</h2> <p>The current solution depends on having a huge address space available. Porting to any 32-bit architecture would quickly run into the limitation of a 2GB or 4GB of address space. The way TM works right now would further divide this limit by N+1, where N is the number of segments. It might be possible to create partially different memory views for multiple threads that each access the same range of addresses; but this would likely require changes inside the OS. We didn't investigate so far.</p> <p>The current 64-bit version relies heavily on Linux- and clang-only features. We believe it is a suitable restriction: a lot of multi- and many-core servers commonly available are nowadays x86-64 machines running Linux. Nevertheless, non-Linux solutions appear to be possible as well. OS X (and likely the various BSDs) seems to handle <tt class="docutils literal">mmap()</tt> better than Linux does, and can remap individual pages of an existing mapping to various pages without hitting a limit of 65536 like Linux. Windows might also have a solution, although we didn't measure yet; but first we would need a 64-bit Windows PyPy, which has not seen much active support.</p> <p>We will likely explore the OS X path (as well as the Windows path if Win64 support grows in PyPy), but this is not part of this current donation proposal.</p> <p>It might be possible to adapt the work done on x86-64 to the 64-bit ARMv8 as well. We have not investigated this so far.</p> </div> <div class="section" id="more-readings"> <h2>More readings</h2> <p>See the <a class="reference external" href="http://pypy.readthedocs.org/en/latest/stm.html">STM page of our documentation</a> and <a class="reference external" href="http://morepypy.blogspot.com/search/label/stm">our blog posts about STM</a>.</p> </div> </div> <div class="section" id="work-plan-and-funding-details"> <span id="work-plan"></span><h1>Work plan and funding details</h1> <p>This is an very rough estimate of the amount of work it would take to complete the steps for an experienced developer who is already familiar with the PyPy codebase. As before, we cannot guarantee the time estimates here, but we do agree to report regularly to the community, so our progress can be followed publicly. We currently expect the duration of the whole project to be up to two years starting from April 2014.</p> <p>Paid work will be at $60/hour, but at least one developer who will work on the project – Armin Rigo – has committed to one hour of volunteer work per paid hour; and another developer – Remi Meier – is a Ph.D. student and gets paid from another source already. The total amount of money that we ask below corresponds roughly to one half-time job.</p> <p>A 10% general donation will go to the <a class="reference external" href="http://sfconservancy.org/">Software Freedom Conservancy</a> itself, the non-profit organization of which the PyPy project is a member and which manages all the issues related to donations, payments, and tax-exempt status. An extra fraction of the money collected will be entered into the general PyPy pot, used for example to finance sprint travel costs to students. This fraction is 10% maximum, unless more money than requested is collected, in which case the whole excess will go to the general PyPy pot.</p> <p><strong>Note</strong> For donations higher than $1,000, we can arrange for an invoice and a different payment method to avoid the high Paypal fees. Please contact pypy at sfconservancy.org if you want to know details on how to donate via other means.</p> <div class="section" id="goal-1"> <h2>Goal 1</h2> <p>The PyPy-TM that we have in the end of March 2014 is good enough in some cases to run existing multithreaded code without a GIL, but not in all of them. There are a number of caveats for the user and missing optimizations. The goal #1 is to improve this case and address the caveats. The current status is written down <a class="reference external" href="http://pypy.readthedocs.org/en/latest/stm.html">in the docs</a> and will evolve over time.</p> <p>For future reference, at the end of March the main identified issues are:</p> <ul class="simple"> <li>There are still a number of bugs.</li> <li>The JIT warm-up time is abysmal.</li> <li>The GC is missing a number of optimizations that are present in a regular PyPy.</li> <li>Destructors are not supported (<tt class="docutils literal">__del__()</tt> methods).</li> <li>The STM bookkeeping logic could see more improvements.</li> <li>Forking the process is slow.</li> <li>We don't foresee particularly high conflict rates in regular multithreaded programs, but this assertion needs to be checked and possibly some heuristics improved.</li> </ul> <p>Fixing all these issues is required before we can confidently say that PyPy-TM is an out-of-the-box replacement of a regular PyPy which gives speed-ups over the regular PyPy independently of the Python program it runs, as long as it is using at least two threads.</p> </div> <div class="section" id="goal-2"> <h2>Goal 2</h2> <p>This goal contains the various new features needed to use multiple cores without explicitly using multithreading; in other words, the new APIs and libraries accessible from Python programs that want to make use of this benefit.</p> <p>This goal requires good support for very-long-running transactions, started with the <tt class="docutils literal">with atomic</tt> construct documented <a class="reference external" href="http://pypy.readthedocs.org/en/latest/stm.html">here</a>. This approach hides the notion of threads from the end programmer, including all the hard multithreading-related issues. This is not the first alternative approach to explicit threads; for example, <a class="reference external" href="http://en.wikipedia.org/wiki/OpenMP">OpenMP</a> is one. However, it is one of the first ones which does not require the code to be organized in a particular fashion. Instead, it works on any Python program which has got latent, imperfect parallelism. Ideally, it only requires that the end programmer identifies where this parallelism is likely to be found, and communicates it to the system, using some lightweight library on top of <tt class="docutils literal">with atomic</tt>.</p> <p>However, this introduces new issues. The main one is that by forcing transactions to be longer, “conflicts” will become more common, up to the point of partially or completely offsetting the benefit of using PyPy-TM in the first place.</p> <p>So the programmer using PyPy-TM needs a way to get feedback about what conflicts we get in these long-running transactions, and where they are produced. A first step will be to implement getting “tracebacks” that point to the places where the most time is lost. This could be later integrated into some “profiler”-like variant where we can navigate the conflicts, either in a live program or based on data logs.</p> <p>Some of these conflicts can be solved by improving PyPy-TM directly. The system works on the granularity of objects and doesn't generate false conflicts, but some conflicts may be regarded as “false” anyway: these involve most importantly the built-in dictionary type, for which we would like accesses and writes using independent keys to be truly independent. Other built-in data structures have a similar issue, like lists: ideally, writes to different indexes should not cause conflicts; but more generally, we would need a mechanism, possibly under the control of the application, to do things like append an item to a list in a “delayed” manner, to avoid conflicts.</p> <p>Similarly, we might need a way to delay some I/O: doing it only at the end of the transaction rather than immediately, in order to prevent the whole transaction from turning inevitable.</p> <p>The goal 2 is thus the development of tools to inspect and fix the causes of conflicts, as well as fixing the ones that are apparent inside PyPy-TM directly.</p> </div> <div class="section" id="goal-3"> <h2>Goal 3</h2> <p>The third goal is to look at some existing event-based frameworks (for example Twisted, Tornado, Stackless, gevent, …) and attempt to make them use threads and atomic sections internally. We would appreciate help and feedback from people more involved in these frameworks, of course.</p> <p>The idea is to apply the techniques described in the <a class="reference internal" href="#goal-2">goal 2</a> until we get a version of framework X which can transparently parallelize the dispatching and execution of multiple events. This might require some slight reorganization of the core in order to split the I/O and the actual logic into separate transactions.</p> </div> <div class="section" id="funding"> <h2>Funding</h2> <p>We forecast that goal 1 and a good chunk of goal 2 should be reached in around 6 months of work. The remaining parts of goal 2 as well as goal 3 are likely to be more open-ended jobs. We will go with a total estimate of two years in order to get a final, well-tested PyPy-STM with stable performance. The amount sought by this fundraising campaign is USD$80'000, corresponding to one half-time job for 16 months (1200 hours at $60/hour plus 10% overhead).</p> </div> </div> <div class="section" id="benefits-of-this-work-to-the-python-community-and-the-general-public"> <h1>Benefits of This Work to the Python Community and the General Public</h1> <p>Python has become one of the most popular dynamic programming languages in the world. Web developers, educators, and scientific programmers alike all value Python because Python code is often more readable and because Python often increases programmer productivity.</p> <p>Traditionally, languages like Python ran more slowly than static, compiled languages; Python developers chose to sacrifice execution speed for ease of programming. The PyPy project created a substantially improved Python language implementation, including a fast Just-in-time (JIT) compiler. The increased execution speed that PyPy provides has attracted many users, who now find their Python code runs between 2 and 50 times faster under PyPy than under the reference implementation written in C.</p> <p>However, in the presence of today's machines with multiple processors, Python progress lags behind. The issue has been described in the introduction: developers that really need to use multiple CPUs are constrained to select and use one of the multi-process solutions that are all in some way or another hacks requiring extra knowledge and efforts to use. The focus of the work described in this proposal is to offer an alternative in the core of the Python language — an alternative that can naturally integrate with the rest of the program. This alternative is implemented in PyPy.</p> <p>PyPy's developers make all PyPy software available to the public without charge, under PyPy's Open Source copyright license, the permissive MIT License. PyPy's license assures that PyPy is equally available to everyone freely on terms that allow both non-commercial and commercial activity. This license allows for academics, for-profit software developers, volunteers and enthusiasts alike to collaborate together to make a better Python implementation for everyone.</p> <p>PyPy-TM is and continues to be available under the same license. Being licensed freely to the general public means that opportunities to use, improve and learn about how Transactional Memory works itself will be generally available to everyone.</p> </div> </div> <div id="sidebar"> </div> </div> </div> </div></div></div> </body> </html>