1 |
Richard Freeman <rich@××××××××××××××.net> posted |
2 |
46B33881.2060801@××××××××××××××.net, excerpted below, on Fri, 03 Aug 2007 |
3 |
10:15:29 -0400: |
4 |
|
5 |
> Actually, as long as you have enough swap your RAM doesn't really |
6 |
> matter. In the worst case it just swaps to disk, which is what it would |
7 |
> be doing anyway if you didn't have a tmpfs. Even if only 10% of the |
8 |
> short-lived files didn't make it to disk it would be a huge savings in |
9 |
> IO waits. |
10 |
|
11 |
Thanks for pointing that out, yes. |
12 |
|
13 |
> Also, consider using ionice (from schedutils I think) - it lets you |
14 |
> schedule IO the way nice schedules CPU. It can have a big impact if you |
15 |
> have a lot of disk churn. |
16 |
|
17 |
It can. However, with the proper i/o scheduler (CFQ), plain nice is |
18 |
often good enough, since the i/o scheduler implements logical defaults |
19 |
based on nice value -- higher nice means higher default ionice. |
20 |
|
21 |
That did get me to switch i/o schedulers here, BTW. I had been using |
22 |
anticipatory until I setup the RAID, at which point I switched to the |
23 |
simpler deadline, since anticipatory does its anticipation stuff assuming |
24 |
a single spindle. After reading a bit about the workings of ionice |
25 |
however, and how CFQ defaults i/o priority based on CPU timeslice |
26 |
priority (and therefore nice), I switched to it. With the memory I have, |
27 |
most disk access after the initial is cached, so it didn't affect me a |
28 |
lot, and RAID helps in that regard as well, but it did seem to make some |
29 |
difference. |
30 |
|
31 |
So anyway, I'm in pretty strong agreement with you. =8^) |
32 |
|
33 |
-- |
34 |
Duncan - List replies preferred. No HTML msgs. |
35 |
"Every nonfree program has a lord, a master -- |
36 |
and if you use the program, he is your master." Richard Stallman |
37 |
|
38 |
-- |
39 |
gentoo-amd64@g.o mailing list |