I trying to get started with Google Perf Tools to profile some CPU intensive applications. It's a statistical calculation that dumps each step to a file using `ofstream'. I'm not a C++ expert so I'm having troubling finding the bottleneck. My first pass gives results:
Total: 857 samples
357 41.7% 41.7% 357 41.7% _write$UNIX2003
134 15.6% 57.3% 134 15.6% _exp$fenv_access_off
109 12.7% 70.0% 276 32.2% scythe::dnorm
103 12.0% 82.0% 103 12.0% _log$fenv_access_off
58 6.8% 88.8% 58 6.8% scythe::const_matrix_forward_iterator::operator*
37 4.3% 93.1% 37 4.3% scythe::matrix_forward_iterator::operator*
15 1.8% 94.9% 47 5.5% std::transform
13 1.5% 96.4% 486 56.7% SliceStep::DoStep
10 1.2% 97.5% 10 1.2% 0x0002726c
5 0.6% 98.1% 5 0.6% 0x000271c7
5 0.6% 98.7% 5 0.6% _write$NOCANCEL$UNIX2003
This is surprising, since all the real calculation occurs in SliceStep::DoStep. The "_write$UNIX2003" (where can I find out what this is?) appears to be coming from writing the output file. Now, what confuses me is that if I comment out all the outfile << "text" statements and run pprof, 95% is in SliceStep::DoStep and `_write$UNIX2003' goes away. However my application does not speed up, as measured by total time. The whole thing speeds up less than 1 percent.
What am I missing?
Added:
The pprof output without the outfile << statements is:
Total: 790 samples
205 25.9% 25.9% 205 25.9% _exp$fenv_access_off
170 21.5% 47.5% 170 21.5% _log$fenv_access_off
162 20.5% 68.0% 437 55.3% scythe::dnorm
83 10.5% 78.5% 83 10.5% scythe::const_matrix_forward_iterator::operator*
70 8.9% 87.3% 70 8.9% scythe::matrix_forward_iterator::operator*
28 3.5% 90.9% 78 9.9% std::transform
26 3.3% 94.2% 26 3.3% 0x00027262
12 1.5% 95.7% 12 1.5% _write$NOCANCEL$UNIX2003
11 1.4% 97.1% 764 96.7% SliceStep::DoStep
9 1.1% 98.2% 9 1.1% 0x00027253
6 0.8% 99.0% 6 0.8% 0x000274a6
This looks like what I'd expect, except I see no visible increase in performance (.1 second on a 10 second calculation). The code is essentially:
ofstream outfile("out.txt");
for loop:
SliceStep::DoStep()
outfile << 'result'
outfile.close()
Update: I timing using boost::timer, starting where the profiler starts and ending where it ends. I do not use threads or anything fancy.
From my comments:
The numbers you get from your profiler say, that the program should be around 40% faster without the print statements.
The runtime, however, stays nearly the same.
Obviously one of the measurements must be wrong. That means you have to do more and better measurements.
First I suggest starting with another easy tool: the time command. This should get you a rough idea where your time is spend.
If the results are still not conclusive you need a better testcase:
Use a larger problem
Do a warmup before measuring. Do some loops and start any measurement afterwards (in the same process).
Tiristan: It's all in user. What I'm doing is pretty simple, I think... Does the fact that the file is open the whole time mean anything?
That means the profiler is wrong.
Printing 100000 lines to the console using python results in something like:
for i in xrange(100000):
print i
To console:
time python print.py
[...]
real 0m2.370s
user 0m0.156s
sys 0m0.232s
Versus:
time python test.py > /dev/null
real 0m0.133s
user 0m0.116s
sys 0m0.008s
My point is:
Your internal measurements and time show you do not gain anything from disabling output. Google Perf Tools says you should. Who's wrong?
_write$UNIX2003 is probably referring to the write POSIX system call, which outputs to the terminal. I/O is very slow compared to almost anything else, so it makes sense that your program is spending a lot of time there if you are writing a fair bit of output.
I'm not sure why your program wouldn't speed up when you remove the output, but I can't really make a guess on only the information you've given. It would be nice to see some of the code, or even the perftools output when the cout statement is removed.
Google perftools collects samples of the call stack, so what you need is to get some visibility into those.
According to the doc, you can display the call graph at statement or address granularity. That should tell you what you need to know.
Related
I've got gperftools installed and collecting data, which looks reasonable so far. I see one node (?) that gets sampled a lot - but I'm interested in the callers to that node - I don't see them? I've tried callgrind/kcachegrind also, I feel like I'm missing something? Here's a snippet of the output when using --text
Total: 1844 samples
573 31.1% 31.1% 573 31.1% US_strcpy
185 10.0% 41.1% 185 10.0% US_strstr
167 9.1% 50.2% 167 9.1% US_strlen
63 3.4% 53.6% 63 3.4% PS_CompressTable
58 3.1% 56.7% 58 3.1% LX_LexInternal
51 2.8% 59.5% 51 2.8% US_CStrEql
47 2.5% 62.0% 47 2.5% 0x40472984
40 2.2% 64.2% 40 2.2% PS_DoSets
38 2.1% 66.3% 38 2.1% LX_ProcessCatRange
So I'm interested in seeing the callers to US_strcpy, but I don't seem to have any? I do get a nice call graph from kcachegrind for 0x40472984 (still trying to match that to a symbol)
There are several ways:
a) pprof --web or kcachgrind will show you callers nicely if it is captured correctly. It is sometimes useful to do pprof --traces (only with github.com/google/pprof version). Which will be somewhat like low-tech method that Mike mentioned above.
b) if the data is really unavailable, you're having problem with stacktrace capturing and/or symbolization. For that, build gperftools with libunwind and build all of your program with debug info.
I'm trying to measure a function's performance by measuring the time for each iteration.
During the process, I found even if I do nothing, the results still vary quite a bit.
e.g.
volatile long count = 0;
for (int i = 0; i < N; ++i) {
measure.begin();
++count;
measure.end();
}
In measure.end(), I measure the time difference and keep an unordered_map to keep track of the time-count.
I've used clock_gettime as well as rdtsc, but there's always about 1% of the data points lie far away from mean, in a 1000 factor.
Here's what the above loop generates:
T: count percentile
18 117563 11.7563%
19 111821 22.9384%
21 201605 43.0989%
22 541095 97.2084%
23 2136 97.422%
24 2783 97.7003%
...
406 1 99.9994%
3678 1 99.9995%
6662 1 99.9996%
17945 1 99.9997%
18148 1 99.9998%
18181 1 99.9999%
22800 1 100%
mean:21
So whether it's ticks or ns, the worst case 22800 is about 1000 times bigger than mean.
I did isolcpus in grub and was running this with taskset. The simple loop almost does nothing, the hash table to do time-count statistics is outside of the time measurements.
What am I missing?
I'm running this on a laptop with ubuntu installed, CPU is Intel(R) Core(TM) i5-2520M CPU # 2.50GHz
Thank you for all the answers.
The main interrupt that I couldn't stop is the local timer interrupt. And it seems new 3.10 kernel would support tickless. I'll try that one.
I'm trying to understand the Callgrind Profile Format. I found the online description
I thought I understood it fairly well until I encountered the 'Extended Example':
events: Instructions
fl=file1.c
fn=main
16 20
cfn=func1
calls=1 50
16 400
cfl=file2.c
cfn=func2
calls=3 20
16 400
fn=func1
51 100
cfl=file2.c
cfn=func2
calls=2 20
51 300
fl=file2.c
fn=func2
20 700
The description reads: One can see that in "main" only code from line 16 is executed where also the other functions are called. Inclusive cost of "main" is 420, which is the sum of self cost 20 and costs spent in the calls.
How can the inclusive cost of 'main' be 420, when the self cost of only func2 is already 700?
OK, the description is wrong: when i paste this example and open it in kcachegrind, indeed it shows a total inclusive cost of 820. That makes sense. Sorry for the noise.
My C++ program is consuming a lot of CPU, and more so as it runs. I used Google Performance Tools to profile CPU usage, and this is what I got:
(pprof) top
Total: 1343 samples
1330 99.0% 99.0% 1330 99.0% 0x0000000801dcb11c
7 0.5% 99.6% 7 0.5% 0x0000000801dcb11e
4 0.3% 99.9% 4 0.3% program::threadWorker
1 0.1% 99.9% 1 0.1% 0x0000000801dcb110
1 0.1% 100.0% 1 0.1% 0x00007fffffffffc0
However, only 1 out of the 5 processes shown here is an actual function name; the rest are addresses. How can I find out what these addresses pertain to? (Of course, I am most interested in the first address shown above)
Edit: This is how I ran the profiler:
env CPUPROFILE=prof.out ./a.out
[kill program]
pprof ./a.out prof.out
Also, I found the root cause by code inspection. But it would still be nice to have the profiler pinpoint the culprit function rather than an address.
Is it possible you haven't specified the executable when loading the results in google-pprof?
I run it as:
$ google-pprof executable /tmp/executable.hprof --text | less
and can see the function names just fine. Or that those methods are in some shared library not in your path when you run google-pprof?
So recently I was playing around with Django on the Jython platform and wanted to see its performance in "production". The site I tested with was just a simple return HttpResponse("Time %.2f" % time.time()) view, so no database involved.
I tried the following two combinations (measurements done with ab -c15 -n500 -k <url>, everything in Ubuntu Server 10.10 on VirtualBox):
J2EE application server (Tomcat/Glassfish), deployed WAR file
I get results like
Requests per second: 143.50 [#/sec] (mean)
[...]
Percentage of the requests served within a certain time (ms)
50% 16
66% 16
75% 16
80% 16
90% 31
95% 31
98% 641
99% 3219
100% 3219 (longest request)
Obviously, the server hangs for a few seconds once in a while, which is not acceptable. I assume it has something to do with reloading Jython because starting the jython shell takes about 3 seconds, too.
AJP serving using patched flup package (+ Apache as frontend)
Note: flup is the package used by manage.py runfcgi, I had to patch it because flup's threading/forking support doesn't seem to work on Jython (-> AJP was the only working method).
Almost the same results here, but sometimes the last 100 requests don't even get answered at all (but server process still alive).
I'm asking this on SO (instead of serverfault) because it's very Django/Jython-specific. Does anyone have experience with deploying Django sites on Jython? Is there maybe another (faster) way to serve the site? Or is it just too early to use Django on the Java platform?
So as nobody replied, I investigated a bit more and it seems like my problem might have to do with VirtualBox. Using different server OSes (Debian Squeeze, Ubuntu Server), I had similar problems. For example, with simple static file serving, I got this result from the Apache web server (on Debian):
> ab -c50 -n1000 http://ip.of.my.vm/some/static/file.css
Requests per second: 91.95 [#/sec] (mean) <--- quite impossible for static serving
[...]
Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 2 22.1 0 688
Processing: 0 206 991.4 31 9188
Waiting: 0 96 401.2 16 3031
Total: 0 208 991.7 31 9203
Percentage of the requests served within a certain time (ms)
50% 31
66% 47
75% 63
80% 78
90% 156
95% 781
98% 844
99% 9141 <--- !!!!!!
100% 9203 (longest request)
This led to the conclusion that (I don't have a conclusion, but) I think the Java reloading might not be the problem here, rather the virtualization. I will try it on a real host and leave this question unanswered till then.
FOLLOWUP
Now I successfully tested a bare-bones Django site (really just the welcome page) using Jython + AJP over TCP/mod_proxy_ajp on Apache (again with patched flup package). This time on a real host (i7 920, 6 GB RAM). The result proved that my above assumption was correct and that I really should never benchmark on a virtual host again. Here's the result for the welcome page:
Document Path: /jython-test/
Document Length: 2059 bytes
Concurrency Level: 40
Time taken for tests: 24.688 seconds
Complete requests: 20000
Failed requests: 0
Write errors: 0
Keep-Alive requests: 0
Total transferred: 43640000 bytes
HTML transferred: 41180000 bytes
Requests per second: 810.11 [#/sec] (mean)
Time per request: 49.376 [ms] (mean)
Time per request: 1.234 [ms] (mean, across all concurrent requests)
Transfer rate: 1726.23 [Kbytes/sec] received
Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 1 1.5 0 20
Processing: 2 49 16.5 44 255
Waiting: 0 48 16.5 44 255
Total: 2 49 16.5 45 256
Percentage of the requests served within a certain time (ms)
50% 45
66% 48
75% 51
80% 53
90% 69
95% 80
98% 90
99% 97
100% 256 (longest request) # <-- no multiple seconds of waiting anymore
Very promising, I would say. The only downside is that the average request time is > 40 ms whereas the development server has a mean of < 3 ms.