xfs
[Top] [All Lists]

Re: Benchmark

To: linux-xfs@xxxxxxxxxxx
Subject: Re: Benchmark
From: Florin Andrei <florin@xxxxxxx>
Date: 11 May 2001 12:05:00 -0700
In-reply-to: <200105111838.f4BIcUt20833@jen.americas.sgi.com>
References: <200105111838.f4BIcUt20833@jen.americas.sgi.com>
Sender: owner-linux-xfs@xxxxxxxxxxx
On 11 May 2001 13:38:30 -0500, Steve Lord wrote:
> > 
> > Is it possible to mount an XFS partition so that deletes are NOT
> > synchronous?
> 
> No, but you can make all the other transactions synchronous too so that the
> deletes do not look as bad when compared to other operations ;-) This is
> the wsync option.

mwaaahahahahahaha!!!

> The only way to experiment with this is to start commenting out
> xfs_trans_set_sync calls in xfs. Can people either have a little

It would be extremely cool and useful to have a mount option like, for
example, "-o delayed_write_tansactions" (you got the idea...).

It's arguable, yes, but sometimes it makes sense to have non-synchr.
write transactions. Since i installed and i'm taking care of Squid on
www-proxy.sgi.com :-) i'm thinking of situations like this one: imagine
a high-performance web proxy, with lots of requests, and the disk caches
full. Now, because the disks are full, you will have lots of removals
among the objects stored there (since new objects are forced into the
cache, and the old ones must die to make room). If delete is slow, the
whole thing will become extremely I/O bound.
Now, Squid is already I/O bound, so sync writes only make things worse.
Right now, the only place where i see ReiserFS as better than XFS is
Squid's cache. With non-sync write transactions, this will not be an
issue anymore.

I'm not sure about that, but i think this is the way Soft Updates from
FreeBSD is doing things. Maybe someone will correct me if i'm wrong.

And not to mention benchmarks. Now that XFS is becoming more and more
popular, people will start to compare it more often with other FSs. The
only way that XFS is clearly slower than others is deleting large trees.
>From the advertisement p.o.v. ;-) it's good to have a simple way to
avoid that.

> If you want to experiment with XFS with a low latency log then try this
> trick:
> 
>       o create a loop device on a file
>       o make an xfs filesystem with an external log on the loop device

Heh... Clever workaround... ;-)

-- 
Florin Andrei

"Remember, son:
if you never try, you never fail" - Homer Simpson


<Prev in Thread] Current Thread [Next in Thread>