1 |
On Thu, Dec 02, 2004 at 10:55:57AM +0900 or thereabouts, Georgi Georgiev wrote: |
2 |
> > all 5 servers in the rsync.gentoo.org rotation are currently running |
3 |
> > rsync out of ram. i think i can safely say they wouldn't be nearly as |
4 |
> > fast as they are, running off a hard disk. |
5 |
> |
6 |
> Do you have some real numbers? I am not trying to doubt you too much. I |
7 |
> am genuinely curious how big the speedup is and if it is worth the |
8 |
> effort. All I wanted to point in my post is that the speedup is probably |
9 |
> not that great, but I'd really like to see benchmark numbers if someone |
10 |
> went to the trouble of doing it. |
11 |
|
12 |
I can't give real comparison numbers, but I can say that we did see |
13 |
dramatic improvements when moving from an IDE drive to a ram drive on the |
14 |
servers in the rsync.g.o rotation. It's a bit hazy, but I seem to recall |
15 |
getting killed with much more than 20 or so simultaneous connections on an |
16 |
IDE-based rsync.g.o server and no more than 30-40 with a SCSI-based one. |
17 |
With the ram-based servers, we've never seen a ceiling. Most of our |
18 |
servers are serving up ~30 simultaneous connections and ~6Mbps on average |
19 |
and their loads never spike much beyond .4 or so. |
20 |
|
21 |
Out of the 5 servers we have in the rotation, we could probably make do |
22 |
with 2 of them. The rest are mainly there for spikes during releases and |
23 |
redundancy in case of failure. |
24 |
|
25 |
--kurt |