On Tue, May 23, 2000 at 08:37:09PM +0100, Stephen C. Tweedie wrote:
> Hi,
>
> On Thu, May 18, 2000 at 01:48:14PM -0700, Brian Pomerantz wrote:
> >
> > I've attached a plot of the performance curve for the Ciprico 7000
> > attached to the QLogic 2200. The dd I have been using aligns the
> > buffer to a page, which as far as I know should work with Stephen
> > Tweedie's raw I/O. Is there anything else that needs to be done to
> > use the raw device once it has been attached to a block device?
>
> Nope, there shouldn't be.
>
> To allow testing of larger blocksizes, I'll add a tunable sysctl
> limit to the 2.3 raw device code to support larger blocksizes. I'm
> curious, though --- does anybody know what the typical blocksize is
> when applications are using raw I/O, as opposed to benchmark tests?
>
I've been running benchmarks to get my "theoretical Max I/O" numbers
as well as find the best fit for my application given the hardware I
have. For the hardware I'm using, I'll be reading and writing in
1-2MB block sizes. I haven't decided what size we'll use in the end,
but it should be around there. If the hardware performed different, I
would use a different size I'm sure.
I'm involved in porting the global file system Petal/Frangipani from
Digital Unix to Linux. The network block server is a user space
program that serves up blocks for the file system and it will be using
the raw device. We provide high performance computing for scientific
computing at Lawrence Livermore National Laboratory. We tell our
users what size "blocks" they should target on a per machine basis (we
have many different production architectures here at LLNL) so they can
tune their apps and I/O libraries accordingly. For HPC here at the
lab, the size tends to be variable based on the hardware at hand and
what the sweet spot for that hardware is. So far there aren't any
large scale machines running Linux that are in production but I'm
hoping that will change in the next couple of years.
BAPper
-
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to [EMAIL PROTECTED]
