[Top] [All Lists]

followup on benchmarks of an xfs embedded system (without rt section)

To: xfs@xxxxxxxxxxx
Subject: followup on benchmarks of an xfs embedded system (without rt section)
From: Jason Newton <nevion@xxxxxxxxx>
Date: Fri, 27 Jul 2012 03:35:40 -0700
Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20120113; h=mime-version:date:message-id:subject:from:to:content-type; bh=O/mQUOX/Ht1QLG2RwgtYNU3FIKmDdagafGvkAny34A8=; b=y6E7BVitBhCgcjnsg32pzq8Keft8aYfmJVaPHxva80yNNGFDeCuhFLhZROqHEZdznk /0GpbIuExv/3xiDchTswEMFBwja/rmkZNPRh7iXTm5V6Eyp3PCKm45l0PC2LI257eQSR wkZziyXUZPo3hyflxNOJF3Q8RVpD6jJOUbMGm7Z3KFh+26y1lxqHBza5FwXyQLObv31d 4cotytMlfiK30jm78eZ1WHab8eV2tL1raGgn7OZ2YfDWnwG0psXQSsaTpcvz6U62DHjn xKmvK9U83YmhohxlxBfeoWpYiem88uXIqr3In5sGh5Vc8MvaCWxmzHkdtGWaS6lyc4A1 T8Jg==
I previously posted on my latencies and that I would update my simulation program a bit more (to cap it to my actual framerate rather than going as fast as throughput allows.

I've updated the simulation and it turns out it drops just as often as before, despite operating at 60fps rather than ~150fps.  I get an ocational dropping of 8 frames which means 400+ms delay in the write calls.   Then things are just fine for a while, until I miss 8 again.  I multithreaded the stream write calls since that's what I do currently in the real program but I disabled openmp and the numbers are all the same with or without.  iotop reports 330-350mB/s of write activity and htop reports 70-92% cpu usage.

[10:30:33.386036000] [6] min: 11ms max: 472ms avg: 14.1086799ms std: 8.1730857ms count = 6648 dropped: 16 transferred: 36.65G totaltime: 333166ms

I've attached my benchmark program but I use alot of boost c++ with a little internal set of libraries... so you can see what I"m doing but it likely won't compile for you.  I'll also mention that boost is a very low overhead (if any) over all the normal system calls one would use (verified by reading sourcecode in use).

So I'll end on the same question:  any ideas on how to squash out that occational 400+ms latency? Dropping down cpu usage? Should I just record to the raid directly without a filesystem?  I guess that could be an interesting test to run tomorrow...


Attachment: iobench.cpp
Description: Text Data

<Prev in Thread] Current Thread [Next in Thread>