Date: Sun, 19 Sep 1999 22:27:04 -0700 (PDT) From: Matthew Dillon <dillon@apollo.backplane.com> To: Greg Lehey <grog@lemis.com> Cc: Poul-Henning Kamp <phk@critter.freebsd.dk>, dg@root.com, cvs-committers@FreeBSD.ORG, cvs-all@FreeBSD.ORG Subject: Re: User block device access (was: cvs commit: src/sys/miscfs/specfs spec_vnops.c src/sys/sys vnode.h src/sys/kern vfs_subr.c) Message-ID: <199909200527.WAA77336@apollo.backplane.com> References: <16748.937762251@critter.freebsd.dk> <199909191806.LAA73255@apollo.backplane.com> <19990920115556.J55065@freebie.lemis.com>
next in thread | previous in thread | raw e-mail | index | archive | help
:
:Poul-Henning has already replied to this with some figures. I can
:confirm what he says; I've been looking a lot at this sort of thing
:with Vinum. That doesn't answer the question "why", however. I
:suspect that there's something funny in buffer management which is
:reducing performance below what it could be. Vinum has a tool which
:can be used to measure device queue length, and I've seen widely
:differing queue lengths when writing to Vinum block devices: they can
:fluctuate between 100 requests and almost none. I can't explain that
:by the way Vinum works; it will just malloc and issue requests. About
:the only place it could block is in malloc, but the actual memory
:usage fluctuates wildly.
:
:I suspect that this is a bug rather than a feature; the question is,
:is it worth following?
:
:Greg
It is, plain and simply, the fact that you are doing one copy (which
is in fact direct DMA) in the case of raw I/O and doing two copies
(one into the cache - direct DMA, and then a copy to the user buffer )
in the case of buffered I/O.
If you apply the mmap() patches found on
http://www.backplane.com/FreeBSD4/ and do an mmap based test, you will
find that the overhead narrows considerably.
There is nothing sinister going on here, buffer copying eats cpu
plain and simple.
The advantage the block device gives you, apart forom the reblocking
is in its ability to cache the data. If you don't take the caching
into account then *OF COURSE* the raw I/O will appear to be faster!
-Matt
Matthew Dillon
<dillon@backplane.com>
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe cvs-all" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199909200527.WAA77336>