On Sun, 16 Apr 2000, James Manning wrote:
> [Gregory Leblanc]
> > > Recovery is a tad simpler with raid1 done at the lower level simply
> > > because none of the md device ever "dies", just one falls
> > > into degraded
> > > and you can skip an mkraid and let normal recovery take over.
> > > Of course,
> > > that leaves the raid1 read balancing algorithm (arguably the
> > > weak point in
> > > the read performance of 0+1 or 1+0) running in two places
> > > instead of one.
> >
> > Could you elaborate a little? Are you talking about the default 0.90 code,
> > or patched with Mika's brilliant patch? Theoretically, RAID1+RAID0 should
> > be extreemly fast for reads, and only a bit slower for writes, assuming that
> > you're not saturating the bus.
>
> Mika's patch is a straightforward one that improves small, random (ie
> seek-heavy) reads well. I haven't seen it (in my experience) improve
> large sequential reads to the point of raid0 (just in my testing), but
It never will it in it's present form. Single large sequential reads are
always satisfied from one disk only. But if you have more than one
concurrent large sequential reads, they will all benefit from my patch.
I have thought about doing something for single large sequential
read performance, ie making it benefit for extra drive(s) but
if someone has an idea how it could be done in right(tm) in raid1 then
please share :) It is not so straightforward as it seems. Without
large buffers it will look good only on paper and all the
real performance is lost in disk seeks.
But i still think that this large sequential read performance
is not highest priority on _raid1_ case. Very few real world
problems need large sequential read performance. Random reads
are what counts in most cases :)
> it's an issue Mika and I have hashed over many other times, and it's
> not worth banging over again on this list.
>
Sorry James but i couldn't resist :)
> Thankfully, it's now a largely moot issue in the cases I need as
> madvise(MADV_SEQUENTIAL) is around so I can get async forward page-in's
> (the main reason I don't care about seq raid1 read perf much anymore,
> and why I added the mmap/madvise code to tiobench)
>
There !
-- Mika <[EMAIL PROTECTED]>
