[Stgt-devel] sg_turs on stgt iscsi drive is very slow

[FUJITA Tomonori]

From: Ming Zhang <blackmagic02881 at gmail.com>
Subject: Re: [Stgt-devel] sg_turs on stgt iscsi drive is very slow
Date: Thu, 28 Dec 2006 11:23:02 -0500

> On Fri, 2006-12-29 at 01:16 +0900, FUJITA Tomonori wrote:
> > From: Ming Zhang <blackmagic02881 at gmail.com>
> > Subject: Re: [Stgt-devel] sg_turs on stgt iscsi drive is very slow
> > Date: Thu, 28 Dec 2006 11:06:04 -0500
> > 
> > > On Sun, 2006-12-24 at 18:07 +0900, FUJITA Tomonori wrote:
> > > > From: FUJITA Tomonori <fujita.tomonori at lab.ntt.co.jp>
> > > > Subject: Re: [Stgt-devel] sg_turs on stgt iscsi drive is very slow
> > > > Date: Thu, 14 Dec 2006 11:41:27 +0900
> > > > 
> > > > > From: Ming Zhang <mingz at ele.uri.edu>
> > > > > Subject: Re: [Stgt-devel] sg_turs on stgt iscsi drive is very slow
> > > > > Date: Tue, 12 Dec 2006 11:13:33 -0500
> > > > > 
> > > > > > same as yesterday. i think that EPOLL should have problem.
> > > > > 
> > > > > I did some performance tests. I wrongly use EPOLL_WAIT or EPOLL_WAIT
> > > > > is much slower than AIO queue (try r616 if you are interested).
> > > > > 
> > > > > I'll try kevent in -mm kernel shortly (probably, after 2.6.20-rc1).
> > > > 
> > > > Seems that it might takes months to have kevent in mainline. So I
> > > > added tweaked the EPOLL_WAIT patch.
> > > > 
> > > > The latest patch (aioepoll-2.6.20-rc2.diff), I've just put into the
> > > > svn repository, should provide much better performances.
> > > 
> > > unfortunately i did not get much better performance at my side.
> > > 
> > > dl 2.6.19, patch -rc2 patch, patch aioepoll-2.6.20-rc2.diff.
> > > enable scsi target support in kernel and disable all kernel debug
> > > option, compile and boot with new kernel.
> > > make KERNELSRC=<rc2> ISCSI=1
> > > export same 4 disk raid0.
> > 
> > Strange. At least, one person told me that it provides good
> > performances.
> > 
> > Have you tried the standard benchmark like disktest?
> 
> no.
> 
> disktest can send outstanding requests while dd is not. so i prefer to
> use dd to check common case first.

People use iSCSI target software with file systems and database
software, which give outstanding requests. So I don't think that dd is
the common case.

I guess that the poor performance for your benchmark software might be
due to Linux AIO problem (see the latest thread in linux-aio). That
patchset might improve the performances.

Anyway, try postmark, disktest, dbench, etc.


> and sg_turs still run slow.

I guess that we need something like kevent for this. But, again, I
don't think that it's not useful from the users' perspective.