Saturday, October 26, 2013

Making faster under PyPy

If you've ever tried to run your programs with under PyPy,
you've probably experienced some incredible slowness. Take this simple

def f():
    return 1

def main():
    i = 10000000
    while i:
        i -= f()


Running time run five times, and looking at the best
run, here's how PyPy 2.1 stacks up against CPython 2.7.5:

Python Time Normalized to CPython
CPython 2.7.5 3.879s 1.0x
PyPy 2.1 53.330s 13.7x slower

Totally ridiculous. I got turned onto this problem because on one of my
projects CPython takes about 1.5 minutes to run our test suite on the build
bot, but PyPy takes 8-10 minutes.

So I sat down to address it. And the results:

Python Time Normalized to CPython
CPython 2.7.5 3.879s 1.0x
PyPy 2.1 53.330s 13.7x slower
PyPy head 1.433s 2.7x faster

Not bad.

Technical details

So how'd we do it? Previously, using sys.settrace() (which
uses under the hood) disabled the JIT. Except it didn't just disable the JIT,
it did it in a particularly insidious way — the JIT had no idea it was being

Instead, every time PyPy discovered that one of your functions was a hotspot,
it would start tracing to observe what the program was doing, and right when it
was about to finish, coverage would run and cause the JIT to abort. Tracing
is a slow process, it makes up for it by generating fast machine code at the
end, but tracing is still incredibly slow. But we never actually got to the
"generate fast machine code" stage. Instead we'd pay all the cost of tracing,
but then we'd abort, and reap none of the benefits.

To fix this, we adjusted some of the heuristics in the JIT, to better show it
how sys.settrace(<tracefunc>) works. Previously the JIT saw it as an opaque
function which gets the frame object, and couldn't tell whether or not it
messed with the frame object. Now we let the JIT look inside the
<tracefunc> function, so it's able to see that isn't
messing with the frame in any weird ways, it's just reading the line number and
file path out of it.

I asked several friends in the VM implementation and research field if they
were aware of any other research into making VMs stay fast when debugging tools
like are running. No one I spoke to was aware of any (but I
didn't do a particularly exhaustive review of the literature, I just tweeted at
a few people), so I'm pleased to say that PyPy is quite possibly the first VM
to work on optimizing code in debugging mode! This is possible because of our
years spent investing in meta-tracing research.

Happy testing,

Wednesday, October 16, 2013

Update on STM

Hi all,

The sprint in London was a lot of fun and very fruitful. In the last update on STM, Armin was working on improving and specializing the automatic barrier placement. There is still a lot to do in that area, but that work is merged now. Specializing and improving barrier placement is still to be done for the JIT.

But that is not all. Right after the sprint, we were able to squeeze the last obvious bugs in the STM-JIT combination. However, the performance was nowhere near to what we want. So until now, we fixed some of the most obvious issues. Many come from RPython erring on the side of caution and e.g. making a transaction inevitable even if that is not strictly necessary, thereby limiting parallelism. Another problem came from increasing counters everytime a guard fails, which caused transactions to conflict on these counter updates. Since these counters do not have to be completely accurate, we update them non-transactionally now with a chance of small errors.

There are still many such performance issues of various complexity left to tackle: we are nowhere near done. So stay tuned or contribute :)


Now, since the JIT is all about performance, we want to at least show you some numbers that are indicative of things to come. Our set of STM benchmarks is very small unfortunately (something you can help us out with), so this is not representative of real-world performance. We tried to minimize the effect of JIT warm-up in the benchmark results.

The machine these benchmarks were executed on has 4 physical cores with Hyper-Threading (8 hardware threads).

Raytracer from stm-benchmarks: Render times in seconds for a 1024x1024 image:

Interpreter Base time: 1 thread 8 threads (speedup)
PyPy-2.1 2.47 2.56 (0.96x)
CPython 81.1 73.4 (1.1x)
PyPy-STM 50.2 10.8 (4.6x)

For comparison, disabling the JIT gives 148s on PyPy-2.1 and 87s on PyPy-STM (with 8 threads).

Richards from PyPy repository on the stmgc-c4 branch: Average time per iteration in milliseconds:

Interpreter Base time: 1 thread 8 threads (speedup)
PyPy-2.1 15.6 15.4 (1.01x)
CPython 239 237 (1.01x)
PyPy-STM 371 116 (3.2x)

For comparison, disabling the JIT gives 492ms on PyPy-2.1 and 538ms on PyPy-STM.

Try it!

All this can be found in the PyPy repository on the stmgc-c4 branch. Try it for yourself, but keep in mind that this is still experimental with a lot of things yet to come. Only Linux x64 is supported right now, but contributions are welcome.

You can download a prebuilt binary from here: (Linux x64 Ubuntu >= 12.04). This was made at revision bafcb0cdff48.


What the numbers tell us is that PyPy-STM is, as expected, the only of the three interpreters where multithreading gives a large improvement in speed. What they also tell us is that, obviously, the result is not good enough yet: it still takes longer on a 8-threaded PyPy-STM than on a regular single-threaded PyPy-2.1. However, as you should know by now, we are good at promising speed and delivering it... years later :-)

But it has been two years already since PyPy-STM started, and this is our first preview of the JIT integration. Expect major improvements soon: with STM, the JIT generates code that is completely suboptimal in many cases (barriers, allocation, and more). Once we improve this, the performance of the STM-JITted code should come much closer to PyPy 2.1.


Remi & Armin

Tuesday, October 15, 2013

Incremental Garbage Collector in PyPy

Hello everyone.

We're pleased to announce that as of today, the default PyPy comes with a GC that has much smaller pauses than yesterday.

Let's start with explaining roughly what GC pauses are. In CPython each object has a reference count, which is incremented each time we create references and decremented each time we forget them. This means that objects are freed each time they become unreachable. That is only half of the story though. First note that when the last reference to a large tree of objects goes away, you have a pause: all the objects are freed. Your program is not progressing at all during this pause, and this pause's duration can be arbitrarily large. This occurs at deterministic times, though. But consider code like this:

class A(object):

a = A()
b = A()
a.item = b
b.item = a
del a
del b

This creates a reference cycle. It means that while we deleted references to a and b from the current scope, they still have a reference count of 1, because they point to each other, even though the whole group has no references from the outside. CPython employs a cyclic garbage collector which is used to find such cycles. It walks over all objects in memory, starting from some known roots, such as type objects, variables on the stack, etc. This solves the problem, but can create noticeable, nondeterministic GC pauses as the heap becomes large and convoluted.

PyPy essentially has only the cycle finder - it does not bother with reference counting, instead it walks alive objects every now and then (this is a big simplification, PyPy's GC is much more complex than this). Although this might sound like a missing feature, it is really one of the reasons why PyPy is so fast, because at the end of the day the total time spent in managing the memory is lower in PyPy than CPython. However, as a result, PyPy also has the problem of GC pauses.

To alleviate this problem, which is essential for applications like games, we started to work on incremental GC, which spreads the walking of objects and cleaning them across the execution time in smaller intervals. The work was sponsored by the Raspberry Pi foundation, started by Andrew Chambers and finished by Armin Rigo and Maciej Fijałkowski.


Everyone loves benchmarks. We did not measure any significant speed difference on our quite extensive benchmark suite on The main benchmark that we used for other comparisons was translating the topaz ruby interpreter using various versions of PyPy and CPython. The exact command was python <pypy-checkout>/bin/rpython -O2 --rtype Versions:

  • topaz - dce3eef7b1910fc5600a4cd0afd6220543104823
  • pypy source - defb5119e3c6
  • pypy compiled with minimark (non-incremental GC) - d1a0c07b6586
  • pypy compiled with incminimark (new, incremental GC) - 417a7117f8d7
  • CPython - 2.7.3

The memory usage of CPython, PyPy with minimark and PyPy with incminimark is shown here. Note that this benchmark is quite bad for PyPy in general, the memory usage is higher and the amount of time taken is longer. This is due to the JIT warmup being both memory hungry and inefficient (see below). But first, the new GC is not worse than the old one.

EDIT:Red line is CPython, blue is incminimark (new), green is minimark (old)

The image was obtained by graphing the output of

However, the GC pauses are significantly smaller. For PyPy the way to get GC pauses is to measure time between start and stop while running stuff with PYPYLOG=gc-collect:log pypy, for CPython, the magic incantation is gc.set_debug(gc.DEBUG_STATS) and parsing the output. For what is worth, the average and total for CPython, as well as the total number of events are not directly comparable since it only shows the cyclic collector, not the reference counts. The only comparable thing is the amount of long pauses and their duration. In the table below, pause duration is sorted into 8 buckets, each meaning "below that or equal to the threshold". The output is generated using the gcanalyze tool.


150.1ms 300.2ms 450.3ms 600.5ms 750.6ms 900.7ms 1050.8ms 1200.9ms
5417 5 3 2 1 1 0 1

PyPy minimark (non-incremental GC):

216.4ms 432.8ms 649.2ms 865.6ms 1082.0ms 1298.4ms 1514.8ms 1731.2ms
27 14 6 4 6 5 3 3

PyPy incminimark (new incremental GC):

15.7ms 31.4ms 47.1ms 62.8ms 78.6ms 94.3ms 110.0ms 125.7ms
25512 122 4 1 0 0 0 2

As we can see, while there is still work to be done (the 100ms ones could be split among several steps), we did improve the situation quite drastically without any actual performance difference.

Note about the benchmark - we know it's a pretty extreme case of JIT warmup, we know we suck on it, we're working on it and we're not afraid of showing PyPy is not always the best ;-)

Nitty gritty details

Here are some nitty gritty details for people really interested in Garbage Collection. This was done as a patch to "minimark", our current GC, and called "incminimark" for now. The former is a generational stop-the-world GC. New objects are allocated "young", which means that they initially live in the "nursery", a special zone of a few MB of memory. When the nursery is full, a "minor collection" step moves the surviving objects out of the nursery. This can be done quickly (a few millisecond) because we only need to walk through the young objects that survive --- usually a small fraction of all young objects; and also by far not all objects that are alive at this point, but only the young ones. However, from time to time this minor collection is followed by a "major collection": in that step, we really need to walk all objects to classify which ones are still alive and which ones are now dead ("marking") and free the memory occupied by the dead ones ("sweeping"). You can read more details here.

This "major collection" is what gives the long GC pauses. To fix this problem we made the GC incremental: instead of running one complete major collection, we split its work into a variable number of pieces and run each piece after every minor collection for a while, until there are no more pieces. The pieces are each doing a fraction of marking, or a fraction of sweeping. It adds some few milliseconds after each of these minor collections, rather than requiring hundreds of milliseconds in one go.

The main issue is that splitting the major collections means that the main program is actually running between the pieces, and so it can change the pointers in the objects to point to other objects. This is not a problem for sweeping: dead objects will remain dead whatever the main program does. However, it is a problem for marking. Let us see why.

In terms of the incremental GC literature, objects are either "white", "gray" or "black". This is called tri-color marking. See for example this blog post about Rubinius, or this page about LuaJIT or the wikipedia description. The objects start as "white" at the beginning of marking; become "gray" when they are found to be alive; and become "black" when they have been fully traversed. Marking proceeds by scanning grey objects for pointers to white objects. The white objects found are turned grey, and the grey objects scanned are turned black. When there are no more grey objects, the marking phase is complete: all remaining white objects are truly unreachable and can be freed (by the following sweeping phase).

In this model, the important part is that a black object can never point to a white object: if the latter remains white until the end, it will be freed, which is incorrect because the black object itself can still be reached. How do we ensure that the main program, running in the middle of marking, will not try to write a pointer to white object into a black object? This requires a "write barrier", i.e. a piece of code that runs every time we set a pointer into an object or array. This piece of code checks if some (hopefully rare) condition is met, and calls a function if that is the case.

The trick we used in PyPy is to consider minor collections as part of the whole, rather than focus only on major collections. The existing minimark GC had always used a write barrier of its own to do its job, like any generational GC. This existing write barrier is used to detect when an old object (outside the nursery) is modified to point to a young object (inside the nursery), which is essential information for minor collections. Actually, although this was the goal, the actual write barrier code is simpler: it just records all old objects into which we write any pointer --- to a young or old object. As we found out over time, doing so is not actually slower, and might actually be a performance improvement: for example, if the main program does a lot of writes into the same old object, we don't need to check over and over again if the written pointer points to a young object or not. We just record the old object in some list the first time, and that's it.

The trick is that this unmodified write barrier works for incminimark too. Imagine that we are in the middle of the marking phase, running the main program. The write barrier will record all old objects that are being modified. Then at the next minor collection, all surviving young objects will be moved out of the nursery. At this point, as we're about to continue running the major collection's marking phase, we simply add to the list of pending gray objects all the objects that we just considered --- both the objects listed as "old objects that are being modified", and the objects that we just moved out of the nursery. A fraction from the former list were black object; so this mean that they are turned back from the black to the gray color. This technique implements nicely, if indirectly, what is called a "backward write barrier" in the literature. The backwardness is about the color that needs to be changed in the opposite of the usual direction "white -> gray -> black", thus making more work for the GC. (This is as opposed to "forward write barrier", where we would also detect "black -> white" writes but turn the white object gray.)

In summary, I realize that this description is less about how we turned minimark into incminimark, and more about how we differ from the standard way of making a GC incremental. What we really had to do to make incminimark was to write logic that says "if the major collection is in the middle of the marking phase, then add this object to the list of gray objects", and put it at a few places throughout minor collection. Then we simply split a major collection into increments, doing marking or sweeping of some (relatively arbitrary) number of objects before returning. That's why, after we found that the existing write barrier would do, it was not much actual work, and could be done without major changes. For example, not a single line from the JIT needed adaptation. All in all it was relatively painless work. ;-)

armin and fijal