Re: [Bug 10846] Slow write on LSISAS1068E (SAS6/iR) on kernel >= 2.6.22

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Tue, 2008-06-03 at 09:09 -0700, bugme-daemon@bugzilla.kernel.org
wrote: 
> Should we consider, that such a difference is normal. You suggest that before
> the kernel was dangerous at writing data?

It's caching too much streaming data, yes.  The danger is largely the
amount of data you lose on a crash and mismanagement of the cache
starving other applications. It's not that much of a problem.  At an
estimated write speed of ~70MB/s your 1GB of data is only around 15s to
effect a full writeout.

> So unpacking linux kernel (for example) is 5x times slower and it is not a
> problem. That's just new way of caching data? We seldom see crashes (thanks to
> linux) and we have power supply.

There are many factors that could account for that.

> I'd like testing on same hardware windows to see its behaviour and speed.
> Anyway thanks for the clarifying of cache: write through. As you say, it was
> not evident. 
> 
> So, now, what can we do to keep 2.6.21 behaviour? That's for a mail system
> writing lots of little files and performance matters too.

Well, initially, I'd try a benchmark that simulates the actual problem,
like postmark.  A streaming write tells you very little about the entire
system performance under a mail server type load.

James