Re: Disk performance under CURRENT

On Mon, 24 May 2004, Kevin Oberman wrote:

> Sorry, but I've been off-line for a while on travel. And my wireless
> card died. :-(
> 
> In any case, by kernel is built without witness, so I get sysctl: 
> unknown oid 'debug.witness_watch'. The system is a UP P4-M and not
> running APIC. No INVARIANTS, either. 

Could you investigate the following possible areas to explore?

(1) Could you see if the ATA configuration differs between the two
    versions, especially relating to DMA mode.  There are substantial ATA
    driver changes, so it could be different parameters are being used.

(2) Could you investigate how the performance difference changes with
    block size?  In particular, I would expect the performance difference
    to be most visible with small block sizes, and least visible with
    large block sizes.  I think before you used a 256k block size, which I
    would classify as on the larger side, but it would be interesting to
    see a table like the following fleshed out:


			4.x	5.x
	1 byte
	256 bytes
	512 bytes
	1k
	2k
	4k
	8k
	16k
	32k
	64k
	128k
	256k
	512k
	1m

    For each, a sustained transfer of at least 30 seconds would be good.

(3) Could you run systat -vmstat, iostat, or some other variation during
    the benchmark and determine the levels of system, interrupt, user, and
    idle CPU during the benchmark.

(4) I don't know much about the machine you're running on, but could you
    try compiling a kernel without SMP and see what difference that makes,
    if any?

Robert N M Watson             FreeBSD Core Team, TrustedBSD Projects
robert_at_fledge.watson.org      Senior Research Scientist, McAfee Research