On Wed, Apr 02, 2003 at 09:44:31PM -0500, Chris Hedemark wrote:
> While the server is admittedly an older machine, for the purpose of
> this test it should not matter as long as the hardware configuration is
> equal for all tests. If we agree on a test suite there is nothing to
That's false.
One of the big problems with a lot of tuning info is that it tends
not to take int consideration hardware, &c. I can tell you for sure
that if you have a giant-cache array connected by fibre channel, _it
makes no difference_ what the filesystem is. The array is so fast
that you can't really fill the cache under normal load anyway.
Similarly, if you have enough memory, every read test is going to be
as fast as any other: you'll get 100% cache hits, and the same memory
configured the same way will always respond at about the same speed.
That said, I think you're right to demand some tests, and to say that
holding the machine constant and changing filesystems is a good
filesystem test.
So here are some suggested things, in no real order:
1. Make sure you run out of buffers before you start to read
(for read filesystem speed tests).
2. Pull the power plug repeatedly while the server is under
load. Judge robustness.
3. Put WAL and data area on different filesystems (to be fair,
this should probably be different spindles, but I'll take what I can
get) and configure the filesystems in various ways (including, say,
writeback for data and full journalling for WAL). See tests above.
4. Make sure your controller doesn't lie about fsync.
5. Test under different loads. 10% writes vs. 90% reads;
20% writes; &c. Compare simple INSERT write with UPDATE write.
Compare UPDATE writes where the UPDATEd row is the same one over and
over. Make sure you do (2) several times.
Lots of these are artificial. But it seems they might reveal
something. I'd be particularly keen to hear about what _really_ is
up with reiserfs.
A
--
----
Andrew Sullivan 204-4141 Yonge Street
Liberty RMS Toronto, Ontario Canada
<andrew@libertyrms.info> M2P 2A8
+1 416 646 3304 x110
