From: bde@runx.oz.au (Bruce Evans) Subject: Re: IDE faster than SCSI-2 Date: Sun, 8 Nov 1992 07:29:24 GMT
In article <1df7dqINN37o@tartarus.uwa.edu.au> oreillym@tartarus.uwa.edu.au (Michael O'Reilly) writes:
>Othman Ahmad (eoahmad@ntuix.ntu.ac.sg) wrote:
>: [...]
>: JUlian's machine
>: is a 50MHz EISA 486 with a 1.3GB drive attached via a Bustek 742a SCSI2
>: adapter. It has 16MB of ram. Part of the disk (about 200MB is taken up
>: by mach2.6 and is unavailable to 386bsd.
>:
>: I checked the load on this machine, it is very light. There is only another
>: user apart from I.
>:
>: iozone 1
>:
>: Writing the 1 Megabyte file, 'iozone.tmp'...1.766667 seconds
> ^
> ^ This is useless. It tells no-one anything. A one meg
>file should fit entirely in the cache. So unless BSD is braindead, and
>doesn't cache (which I doubt), then you have just measured the cache
>speed. On my 486/33, 8 meg ram, a 1 meg files get 10 megs per second.
1 MB in 1.76 seconds is way below the cache speed. In fact it is quite
likely to be a good estimate for actual disk speed.
Disk caching in 386BSD-0.1 is quite broken. First, only 10% of memory is
used for the disk+swap cache. This is smaller than I like. (Comments in
the code say that only 5% of memory after the first 2MB are allocated to
the cache, but due to bugs, 10% of all memory is allocated. A fix has
been posted.) Second, only half of the 10% is used for the disk cache.
(Some (usually all) of the other half is reserved for buffers for swapping.)
Third, memory for each buffer is allocated twice, so only 2.5% of memory is
used for the disk cache! A fix is in Terry's patch kit. Fourth, bugs
in the caching algorithm caused the cache to thrash. A fix is in the patch
kit. Fifth, another kernel limit requires limiting the initial 10%
allocation to 256 buffers to avoid panics (the limit is in the patch kit).
(I actually use a limit of 503 to get 251 disk buffers and 252 swap buffers.
This is fairly safe because the system almost never swaps.)
>I don't know exactly what you measured, but it sure ain't disk speed.
It probably was disk speed! I don't know exactly what the cache size
is on Julian's system. If the system has all the standard patches, it
will have 128 disk buffers for a total cache size of about 512K if the
file system has a block size of 4K. That is consistent with the
benchmark.
-- Bruce Evans (bde@runx.oz.au)