Wireshark-dev: Re: [Wireshark-dev] The cost of memory allocation
From: Anders Broman <[email protected]>
Date: Wed, 21 Sep 2016 12:07:24 +0000


Just briefly browsing the code…

Could proto_get_finfo_ptr_array() be used instead of proto_find_finfo()?

Perhaps these functions should be rewritten to use wmem arrays instead or use g_ptr_array_sized_new ()




From: [email protected] [mailto:wireshark-dev-[email protected]] On Behalf Of Graham Bloice
Sent: den 21 september 2016 11:15
To: Developer support list for Wireshark <[email protected]>
Subject: Re: [Wireshark-dev] The cost of memory allocation




On 21 September 2016 at 10:12, Graham Bloice <[email protected]> wrote:



On 21 September 2016 at 10:06, Paul Offord <[email protected]> wrote:

Good point – debug build.


Debug builds using the ms allocator are a lot slower due to all the extra memory checking, i.e. see this page: https://msdn.microsoft.com/en-us/library/974tc9t1.aspx


I'm not entirely certain though, that a debug build of Wireshark will use a version of glib that then uses the debug calls into msvcrt.


However the point still stands, that using debug builds for performance testing might not be giving you the real picture.


And this post explains how even a release build is affected by debug memory allocations if it's run under a debugger: http://preshing.com/20110717/the-windows-heap-is-slow-when-launched-from-the-debugger/




From: [email protected] [mailto:[email protected]] On Behalf Of Graham Bloice
Sent: 21 September 2016 09:49
To: Developer support list for Wireshark <[email protected]>
Subject: Re: [Wireshark-dev] The cost of memory allocation




On 21 September 2016 at 09:29, Paul Offord <[email protected]> wrote:

I’m not happy with the performance of the transum dissector and so I’ve started some analysis.  I’ve never used VS performance profiling before but I plan try to investigate this problem using it.  In the meantime I’ve used a tool that I’m reasonably familiar with called PerfView.  It’s produced some interesting results which I thought I’d share.


The problem I’m having is that with transum enabled load time for a 50MB file increases from 5 seconds to 10 seconds, but then subsequent loads of the same file go out to about 40 or 50 seconds.



Above (or attached depending on your email system) is a screen shot showing the time spent in various functions when a load of the file took 44.8 seconds.  At the top of the image is a transum function called decode_gtcp.  The image shows that 50.7% of the total load time was spent executing in this function.  Then we see all of the nested functions with the proportion of time spent in each of those.


What I notice is that a lot of time is being spent in glib functions, and in particular the time is being spent allocating and freeing memory.



Using a slightly different view we can see that across the whole of the process during a load file with transum enabled more than 66% of the time is spent messing around with memory.


I haven’t yet figured out why I get inconsistent load times, and I don’t know what I can do about any of the above, but I thought it might be of general interest.


Best regards…Paul



Release or debug build?



Graham Bloice




Graham Bloice




Graham Bloice