memory vs performance [LWN.net]

memory vs performance

Posted Apr 5, 2012 8:34 UTC (Thu) by dlang (guest, #313)
Parent article: 2012 Linux Storage, Filesystem, and Memory Management Summit - Day 1

This is a really hard problem to identify, on many systems there is no thrashing to indicate that more memory will help.

I recently upgraded the memory on a machine based on my judgement that it would help (and it did, a nightly processing run went from 10+ hours to <2 hours), but this system had no paging taking place, the only performance stats that changed were disk utilization (and I/O wait) and the fact that the CPU utilization shot up.

I've run into this same sort of situation multiple times where the real underlying cause is that the disk cache no longer holds the working set and so disk I/O skyrockets and performance crashes. This is sometimes complicated that 'working set' may require looking at a very long timeframe. In my most recent situation it was analysing log files for the day, so data untouched for the last 24 hours was still part of the overall 'working set'

This is a very hard problem

to post comments

memory vs performance

Posted Apr 5, 2012 12:45 UTC (Thu) by alankila (guest, #47141) [Link] (3 responses)

Perhaps the problem could be diagnosed by observing the read/write ratio. A fully cached scheme results in no read I/O whatsoever, while uncached case could issue far more reads than writes. In any case, one has to form some kind of estimate to determine what the acceptable read/write ratio should be for that particular job.

memory vs performance

Posted Apr 5, 2012 20:21 UTC (Thu) by dlang (guest, #313) [Link] (2 responses)

the problem is how can the system tell if the reads are needed because the data hasn't been available, or if they are wasteful because the data was available in the past (and then the question is how far in the past, in my case it was 24 hours + processing time, which is an eternity to the system)

memory vs performance

Posted Apr 6, 2012 18:26 UTC (Fri) by alankila (guest, #47141) [Link] (1 responses)

I'm not suggesting this is an automatic scheme. I doubt it is even possible to design. I'm just suggesting that if you have an understanding of the behavior you can use this sort of metric to determine if the system is behaving poorly or well.

memory vs performance

Posted Apr 6, 2012 19:39 UTC (Fri) by dlang (guest, #313) [Link]

ahh, I misunderstood you. Yes, watching for this sort of thing is part of tuning the system.