pypy / pypy / tool / memusage /

#! /usr/bin/env python
Produces a logfile.gnumeric file which contains the data extracted from the
logfile generated with the PYPYLOG env variable.

Run your program like this::

    $ PYPYLOG=gc-collect,jit-mem:logfile pypy

This will produce "logfile", containing information about the memory used by
the GC and the number of loops created/freed by the JIT.

If you want, you can also measure the amout of used memory as seen by the OS
(the VmRSS) using

    $ PYPYLOG=gc-collect,jit-mem:logfile ./ -o logfile.vmrss /path/to/pypy

log2gnumeric will automatically pick logfile.vmrss, if present.

If you want to compare PyPy to CPython, you can add its VmRSS to the graph, by
using the -c option.  To produce the .vmrss file, use again ./

    $ ./ -o cpython.vmrss python
    $ ./ -c cpython.vmrss logfile

Note that on CPython it will take a different amout of time to complete, but
on the graph the plot will be scaled to match the duration of the PyPy run
(i.e., the two lines will end "at the same time").

If you are benchmarking, you can add the "translation-task"
category to the log, by setting PYPYLOG=gc-collect,jit-mem,translation-task.

You can freely edit the graph in log-template.gnumeric: this script will
create a new file replacing the 'translation-task' and 'gc-collect' sheets.

import re, sys
import gzip
import optparse

def main(logname, options):
    outname = logname + '.gnumeric'
    data = open(logname).read()
    data = data.replace('\n', '')
    minclock, maxclock = get_clock_range(data)
    time0 = minclock # we want "relative clocks"
    maxtime = maxclock-time0
    xml ='log-template.gnumeric').read()
    xml = replace_sheet(xml, 'translation-task', tasks_rows(time0, data))
    xml = replace_sheet(xml, 'gc-collect', gc_collect_rows(time0, data))
    xml = replace_sheet(xml, 'loops', loops_rows(time0, data))
    xml = replace_sheet(xml, 'vmrss', vmrss_rows(logname + '.vmrss', maxtime))
    xml = replace_sheet(xml, 'cpython-vmrss', vmrss_rows(options.cpython_vmrss, maxtime))
    out =, 'wb')

# ========================================================================
# functions to manipulate gnumeric files
# ========================================================================

def replace_sheet(xml, sheet_name, data):
    pattern = '<gnm:Sheet .*?<gnm:Name>%s</gnm:Name>.*?(<gnm:Cells>.*?</gnm:Cells>)'
    regex = re.compile(pattern % sheet_name, re.DOTALL)
    cells = gen_cells(data)
    match =
    if not match:
        print 'Cannot find sheet %s' % sheet_name
        return xml
    a, b = match.span(1)
    xml2 = xml[:a] + cells + xml[b:]
    return xml2

def gen_cells(data):
    # values for the ValueType attribute
    ValueType_Empty  = 'ValueType="10"'
    ValueType_Number = 'ValueType="40"'
    ValueType_String = 'ValueType="60"'
    parts = []
    for i, row in enumerate(data):
        for j, val in enumerate(row):
            if val is None:
                attr = ValueType_Empty
                val = ''
            elif isinstance(val, (int, long, float)):
                attr = ValueType_Number
                attr = ValueType_String
            cell = '        <gnm:Cell Row="%d" Col="%d" %s>%s</gnm:Cell>'
            parts.append(cell % (i, j, attr, val))
    parts.append('      </gnm:Cells>')
    return '\n'.join(parts)

# ========================================================================
# functions to extract various data from the logs
# ========================================================================

def read_clock(x):
    timestamp = int(x, 16)
    return timestamp / CLOCK_FACTOR

def get_clock_range(data):
    s = r"\[([0-9a-f]+)\] "
    r = re.compile(s)
    clocks = [read_clock(x) for x in r.findall(data)]
    return min(clocks), max(clocks)

def gc_collect_rows(time0, data):
    s = r"""
----------- Full collection ------------------
\| used before collection:
\|          in ArenaCollection:      (\d+) bytes
\|          raw_malloced:            (\d+) bytes
\| used after collection:
\|          in ArenaCollection:      (\d+) bytes
\|          raw_malloced:            (\d+) bytes
\| number of major collects:         (\d+)
\[([0-9a-f]+)\] gc-collect\}"""
    r = re.compile(s.replace('\n', ''))
    yield 'clock', 'gc-before', 'gc-after'
    for a,b,c,d,e,f in r.findall(data):
        clock = read_clock(f) - time0
        yield clock, int(a)+int(b), int(c)+int(d)

def tasks_rows(time0, data):
    s = r"""
\[([0-9a-f]+)\] \{translation-task
starting ([\w-]+)
    r = re.compile(s.replace('\n', ''))
    yield 'clock', None, 'task'
    for a,b in r.findall(data):
        clock = read_clock(a) - time0
        yield clock, 1, b

def loops_rows(time0, data):
    s = r"""
\[([0-9a-f]+)\] \{jit-mem-looptoken-(alloc|free)
    r = re.compile(s.replace('\n', ''))
    yield 'clock', 'total', 'loops', 'bridges'
    loops = 0
    bridges = 0
    fake_total = 0
    for clock, action, text in r.findall(data):
        clock = read_clock(clock) - time0
        if text.startswith('allocating Loop #'):
            loops += 1
        elif text.startswith('allocating Bridge #'):
            bridges += 1
        elif text.startswith('freeing Loop #'):
            match = re.match('freeing Loop # .* with ([0-9]*) attached bridges', text)
            loops -=1
            bridges -= int(
        total = loops+bridges
        yield clock, loops+bridges, loops, bridges

def vmrss_rows(filename, maxtime):
    lines = []
    if filename:
            lines = open(filename).readlines()
        except IOError:
            print 'Warning: cannot find file %s, skipping this sheet' % filename
    for row in vmrss_rows_impl(lines, maxtime):
        yield row

def vmrss_rows_impl(lines, maxtime):
    yield 'inferred clock', 'VmRSS'
    numlines = len(lines)
    for i, line in enumerate(lines):
        mem = int(line)
        clock = maxtime * i // (numlines-1)
        yield clock, mem

if __name__ == '__main__':
    CLOCK_FACTOR = 1000000000.0 # report GigaTicks instead of Ticks
    parser = optparse.OptionParser(usage="%prog logfile [options]")
    parser.format_description = lambda fmt: __doc__
    parser.description = __doc__
    parser.add_option('-c', '--cpython-vmrss', dest='cpython_vmrss', default=None, metavar='FILE', type=str,
                      help='the .vmrss file produced by CPython')

    options, args = parser.parse_args()
    if len(args) != 1:
    main(args[0], options)