Re: How does one size a HN? Is RAM or CPU more important? [message #10544 is a reply to message #10538] |
Wed, 21 February 2007 11:27   |
rickb
Messages: 368 Registered: October 2006
|
Senior Member |
|
|
Thank you for the information, truly a valuable post. I have some further questions if you have time.
-------------
What sector readahead value are you using with blockdev? Default on my card is 256 (blockdev --getra /dev/sdX), the pdf suggest 16384 secs.
I am using ext3 but am not using any tuned values beyond noatime. Have you found and performance gains from optimizations on the fs level?
I am using the cfq io scheduler, because after reading and unofficially testing it, it seems to deliver the best io response time for a shared resource system. Although deadline may offer better overall io performance, fairness is more relevant in a shared resource environment. What are your thoughts on this?
My servers with this card have 8GB of ram, and I have created 8GB of swap; after 63 days of uptime, the system stays constant with ~2.5GB of swap utilized. I have read that adjusting the amount of swap in the system can positively effect performance. Do you have any ideas there?
[root@gallium ~]# free -m
total used free shared buffers cached
Mem: 8029 8004 24 0 97 1586
-/+ buffers/cache: 6320 1709
Swap: 8000 2609 5391
My primary indicator of "disk performance" in a shared resource environment is a (lower) iowait value over time (vmstat, sar, iostat, etc). Do you find this relevant?
-Rick Blundell
-------------
Common Terms I post with: http://wiki.openvz.org/Category:Definitions
UBC. Learn it, love it, live it: http://wiki.openvz.org/Proc/user_beancounters
[Updated on: Thu, 22 February 2007 02:50] Report message to a moderator
|
|
|