The effect of TCmalloc in the QtWebKit port - stage 2: Memory consumption

Posted by zoltan.horvath at 10/28/2009 - 14:18
memory
tcmalloc

As I mentioned, there is a reverse of a medal, so now let's see how enabling TCmalloc on the QtWebKit port's JavaScriptCore engine effects memory consumption.

Results for x86

We can see the memory consumption results of QtWebKit on x86-Linux (with JIT) in kilobytes in the following table:

QtWebKit x86-Linux	System malloc	TCmalloc	Difference
SunSpider	12 162 kB	15 209 kB	~25.05% hungrier
V8	139 513 kB	140 270 kB	~0.54% hungrier
WindScorpion	34 046 kB	40 528 kB	~19.03% hungrier

SunSpider
12 162 kB -> 15 209 kB
+ 3 047 kB
~25.05% hungrier

V8
139 513 kB -> 140 270 kB
+ 757 kB
~0.54% hungrier

WindScorpion
34 046 kB -> 40 528 kB
+ 6 482 kB
~19.03% hungrier

Conclusion

As we can expect, TCmalloc eats more memory on all of the three benchmark suites than the system malloc. We have to pay the price for the achieved significant performance improvement with enabling it.

Future

These results represents the memory consumption of the JavaScriptCore part of the WebKit only, so in the future maybe we should investigate the results for the whole WebKit. Moreover, we have several possibilities to try decreasing the memory consumption: tweaking the parameters of TCmalloc, turning off TCmalloc :-) and trying out other allocators.

zecke (not verified) - 11/01/2009 - 08:32

You should describe how you are measuring memory usage and define "eating" memory.

zoltan.horvath - 11/03/2009 - 09:06

We measure the memory consumption (maximum resident set size) through the kernel's getrusage struct (we've a modified Linux kernel). This solution is similar to using the FreeBSD's /usr/bin/time command or the getrusage struct.

Web Apocrypha

Search

Tags

Links