Subject: Re: Disk performance under NetBSD on A3000
To: Eduardo E. Horvath firstname.lastname@example.org <email@example.com>
From: Andreas Johansson <firstname.lastname@example.org>
Date: 04/25/1995 02:18:40
According to Eduardo E. Horvath email@example.com:
> On Tue, 25 Apr 1995, Andreas Johansson wrote:
> > Not my experience with my system. Write: 550kB/s, read 650kB/s.
> > System described in my last letter.
> Raw reads of 650kB/s are awful slow. Your write rate seems about right tho.
Well, better values with 0ms rotational delay:
Write 874kB/s, read 1310kB/s.
> > > 2) You are writing to a file within the filesystem, which involves such
> > > things as file-system overhead, copying from your process space into the
> > > buffer cache, and updating the filesystem structures.
> > Yep, but couldn't this somehow be done more efficiently, the A3640 in my
> > system isn't that quick in moving data...
> > Under ados I have tried a memorymoverprogram.
> > Reads: 12.5Mb/s
> > Writes: 6.?Mb/s (due to copyback cache).
> > Moving data like this isn't exactly improving disk performance.
> > Some speed could be won by using the move16 instruction (with it's alignment
> > problems (?)) or disabling the datacache while moving with a 040. Is
> > anything done to improve speed on a 3640?
> > Anyone?
> If you want fast I/O you should DMA directly to/from disk. Anything else
> is a waste.
Naturally I use dma-transfers, I was talking about your above statement
about copying from userspace to buffer cache. A 25MHz 030 would be 4 times
quicker at writing if copyback was used on the 25MHz 040, and twice as fast
at reading, all due to the 3640 board being adapted to a memory system
developed for the 030. I belive this could have been done better, but now
things are as they are.
> > > 3) You are writing 1K at a time, where the filesystem is using 8K
> > > blocks. This means you will probably be doing 4 writes/block.
> > You are misreading the dd command, it says bs=1024k <---.
I suppose you noticed this?
> > This figure probably wouldn't change much with a slower disk - so why
> > should I have a fast disk with NetBSD?
> Swap, which doesn't use a filesystem.
Yep, that's right. Now the filesystem speed is also doubled so the above
argument is a little out of line.
> > BTW: I haven't tried the tunefs thing, I haven't been able to see the
> > current value either.
> I have, and I don't notice much difference. The FS and COW overhead swamps
> just about anything else.
As I said: I got much better values with 0 ms rotational delay. Perhaps this
is a disk to disk issue.
: E-mail: firstname.lastname@example.org Amiga 4000 is it! :
: S-mail: Andreas Johansson, Karhusvagen 5 6:618, 977 54 Lulea, SWEDEN :
: There are two groups of people I'll _never_ understand: :
: math-professors and women :