On Mon, Mar 17, 2014 at 3:20 PM, Maciej Fijalkowski <fijall@gmail.com> wrote:
are you *sure* it's the walkroots that take that long and not something else (like gc-minor)? More of those mean that you allocate a lot more surviving objects. Can you do two things:
a) take a max of gc-minor (and gc-minor-stackwalk), per request b) take the sum of those
and plot them
^^^ or just paste the results actually
On Mon, Mar 17, 2014 at 3:18 PM, Martin Koch <mak@issuu.com> wrote:
Well, then it works out to around 2.5GHz, which seems reasonable. But it doesn't alter the conclusion from the previous email: The slow queries then all have a duration around 34*10^9 units, 'normal' queries 1*10^9 units, or .4 seconds at this conversion. Also, the log shows that a slow query performs many more gc-minor operations than a 'normal' one: 9600 gc-collect-step/gc-minor/gc-minor-walkroots operations vs 58.
So the question becomes: Why do we get this large spike in gc-minor-walkroots, and, in particular, is there any way to avoid it :) ?
Thanks, /Martin
On Mon, Mar 17, 2014 at 1:53 PM, Maciej Fijalkowski <fijall@gmail.com> wrote:
I think it's the cycles of your CPU
On Mon, Mar 17, 2014 at 2:48 PM, Martin Koch <mak@issuu.com> wrote:
What is the unit? Perhaps I'm being thick here, but I can't correlate it with seconds (which the program does print out). Slow runs are around 13 seconds, but are around 34*10^9(dec), 0x800000000 timestamp units (e.g. from 0x2b994c9d31889c to 0x2b9944ab8c4f49).
On Mon, Mar 17, 2014 at 12:09 PM, Maciej Fijalkowski <fijall@gmail.com> wrote:
The number of lines is nonsense. This is a timestamp in hex.
On Mon, Mar 17, 2014 at 12:46 PM, Martin Koch <mak@issuu.com> wrote:
Based On Maciej's suggestion, I tried the following
PYPYLOG=- pypy mem.py 10000000 > out
This generates a logfile which looks something like this
start--> [2b99f1981b527e] {gc-minor [2b99f1981ba680] {gc-minor-walkroots [2b99f1981c2e02] gc-minor-walkroots} [2b99f19890d750] gc-minor} [snip] ... <--stop
It turns out that the culprit is a lot of MINOR collections.
I base this on the following observations:
I can't understand the format of the timestamp on each logline (the "[2b99f1981b527e]"). From what I can see in the code, this should be output from time.clock(), but that doesn't return a number like that when I run pypy interactively Instead, I count the number of debug lines between start--> and the corresponding <--stop. Most runs have a few hundred lines of output between start/stop All slow runs have very close to 57800 lines out output between start/stop One such sample does 9609 gc-collect-step operations, 9647 gc-minor operations, and 9647 gc-minor-walkroots operations.
Thanks, /Martin
On Mon, Mar 17, 2014 at 8:21 AM, Maciej Fijalkowski <fijall@gmail.com> wrote: > > there is an environment variable PYPYLOG=gc:- (where - is stdout) > which will do that for you btw. > > maybe you can find out what's that using profiling or valgrind? > > On Sun, Mar 16, 2014 at 11:34 PM, Martin Koch <mak@issuu.com> wrote: > > I have tried getting the pypy source and building my own version > > of > > pypy. I > > have modified > > rpython/memory/gc/incminimark.py:major_collection_step() > > to > > print out when it starts and when it stops. Apparently, the slow > > queries > > do > > NOT occur during major_collection_step; at least, I have not > > observed > > major > > step output during a query execution. So, apparently, something > > else > > is > > blocking. This could be another aspect of the GC, but it could > > also > > be > > anything else. > > > > Just to be sure, I have tried running the same application in > > python > > with > > garbage collection disabled. I don't see the problem there, so it > > is > > somehow > > related to either GC or the runtime somehow. > > > > Cheers, > > /Martin > > > > > > On Fri, Mar 14, 2014 at 4:19 PM, Martin Koch <mak@issuu.com> > > wrote: > >> > >> We have hacked up a small sample that seems to exhibit the same > >> issue. > >> > >> We basically generate a linked list of objects. To increase > >> connectedness, > >> elements in the list hold references (dummy_links) to 10 randomly > >> chosen > >> previous elements in the list. > >> > >> We then time a function that traverses 50000 elements from the > >> list > >> from a > >> random start point. If the traversal reaches the end of the list, > >> we > >> instead > >> traverse one of the dummy links. Thus, exactly 50K elements are > >> traversed > >> every time. To generate some garbage, we build a list holding the > >> traversed > >> elements and a dummy list of characters. > >> > >> Timings for the last 100 runs are stored in a circular buffer. If > >> the > >> elapsed time for the last run is more than twice the average > >> time, > >> we > >> print > >> out a line with the elapsed time, the threshold, and the 90% > >> runtime > >> (we > >> would like to see that the mean runtime does not increase with > >> the > >> number of > >> elements in the list, but that the max time does increase > >> (linearly > >> with the > >> number of object, i guess); traversing 50K elements should be > >> independent of > >> the memory size). > >> > >> We have tried monitoring memory consumption by external > >> inspection, > >> but > >> cannot consistently verify that memory is deallocated at the same > >> time > >> that > >> we see slow requests. Perhaps the pypy runtime doesn't always > >> return > >> freed > >> pages back to the OS? > >> > >> Using top, we observe that 10M elements allocates around 17GB > >> after > >> building, 20M elements 26GB, 30M elements 28GB (and grows to 35GB > >> shortly > >> after building). > >> > >> Here is output from a few runs with different number of elements: > >> > >> > >> pypy mem.py 10000000 > >> start build > >> end build 84.142424 > >> that took a long time elapsed: 13.230586 slow_threshold: > >> 1.495401 > >> 90th_quantile_runtime: 0.421558 > >> that took a long time elapsed: 13.016531 slow_threshold: > >> 1.488160 > >> 90th_quantile_runtime: 0.423441 > >> that took a long time elapsed: 13.032537 slow_threshold: > >> 1.474563 > >> 90th_quantile_runtime: 0.419817 > >> > >> pypy mem.py 20000000 > >> start build > >> end build 180.823105 > >> that took a long time elapsed: 27.346064 slow_threshold: > >> 2.295146 > >> 90th_quantile_runtime: 0.434726 > >> that took a long time elapsed: 26.028852 slow_threshold: > >> 2.283927 > >> 90th_quantile_runtime: 0.374190 > >> that took a long time elapsed: 25.432279 slow_threshold: > >> 2.279631 > >> 90th_quantile_runtime: 0.371502 > >> > >> pypy mem.py 30000000 > >> start build > >> end build 276.217811 > >> that took a long time elapsed: 40.993855 slow_threshold: > >> 3.188464 > >> 90th_quantile_runtime: 0.459891 > >> that took a long time elapsed: 41.693553 slow_threshold: > >> 3.183003 > >> 90th_quantile_runtime: 0.393654 > >> that took a long time elapsed: 39.679769 slow_threshold: > >> 3.190782 > >> 90th_quantile_runtime: 0.393677 > >> that took a long time elapsed: 43.573411 slow_threshold: > >> 3.239637 > >> 90th_quantile_runtime: 0.393654 > >> > >> Code below > >> -------------------------------------------------------------- > >> import time > >> from random import randint, choice > >> import sys > >> > >> > >> allElems = {} > >> > >> class Node: > >> def __init__(self, v_): > >> self.v = v_ > >> self.next = None > >> self.dummy_data = [randint(0,100) > >> for _ in xrange(randint(50,100))] > >> allElems[self.v] = self > >> if self.v > 0: > >> self.dummy_links = [allElems[randint(0, self.v-1)] > >> for _ > >> in > >> xrange(10)] > >> else: > >> self.dummy_links = [self] > >> > >> def set_next(self, l): > >> self.next = l > >> > >> > >> def follow(node): > >> acc = [] > >> count = 0 > >> cur = node > >> assert node.v is not None > >> assert cur is not None > >> while count < 50000: > >> # return a value; generate some garbage > >> acc.append((cur.v, [choice("abcdefghijklmnopqrstuvwxyz") > >> for > >> x > >> in > >> xrange(100)])) > >> > >> # if we have reached the end, chose a random link > >> cur = choice(cur.dummy_links) if cur.next is None else > >> cur.next > >> count += 1 > >> > >> return acc > >> > >> > >> def build(num_elems): > >> start = time.time() > >> print "start build" > >> root = Node(0) > >> cur = root > >> for x in xrange(1, num_elems): > >> e = Node(x) > >> cur.next = e > >> cur = e > >> print "end build %f" % (time.time() - start) > >> return root > >> > >> > >> num_timings = 100 > >> if __name__ == "__main__": > >> num_elems = int(sys.argv[1]) > >> build(num_elems) > >> total = 0 > >> timings = [0.0] * num_timings # run times for the last > >> num_timings > >> runs > >> i = 0 > >> beginning = time.time() > >> while time.time() - beginning < 600: > >> start = time.time() > >> elem = allElems[randint(0, num_elems - 1)] > >> assert(elem is not None) > >> > >> lst = follow(elem) > >> > >> total += choice(lst)[0] # use the return value for > >> something > >> > >> end = time.time() > >> > >> elapsed = end-start > >> timings[i % num_timings] = elapsed > >> if (i > num_timings): > >> slow_time = 2 * sum(timings)/num_timings # slow > >> defined > >> as > >> > > >> 2*avg run time > >> if (elapsed > slow_time): > >> print "that took a long time elapsed: %f > >> slow_threshold: > >> %f 90th_quantile_runtime: %f" % \ > >> (elapsed, slow_time, > >> sorted(timings)[int(num_timings*.9)]) > >> i += 1 > >> print total > >> > >> > >> > >> > >> > >> On Thu, Mar 13, 2014 at 7:45 PM, Maciej Fijalkowski > >> <fijall@gmail.com> > >> wrote: > >>> > >>> On Thu, Mar 13, 2014 at 1:45 PM, Martin Koch <mak@issuu.com> > >>> wrote: > >>> > Hi Armin, Maciej > >>> > > >>> > Thanks for responding. > >>> > > >>> > I'm in the process of trying to determine what (if any) of the > >>> > code > >>> > I'm > >>> > in a > >>> > position to share, and I'll get back to you. > >>> > > >>> > Allowing hinting to the GC would be good. Even better would be > >>> > a > >>> > means > >>> > to > >>> > allow me to (transparently) allocate objects in unmanaged > >>> > memory, > >>> > but I > >>> > would expect that to be a tall order :) > >>> > > >>> > Thanks, > >>> > /Martin > >>> > >>> Hi Martin. > >>> > >>> Note that in case you want us to do the work of isolating the > >>> problem, > >>> we do offer paid support to do that (then we can sign NDAs and > >>> stuff). > >>> Otherwise we would be more than happy to fix bugs once you > >>> isolate > >>> a > >>> part you can share freely :) > >> > >> > >