PROJECT: raw throughoutput in kernel

Marc SCHAEFER (schaefer@alphanet.ch)
10 Jun 1998 10:14:23 +0200


To evaluate raw disk performance, we currently use MS-WINDOWS'NT
with either the SCSI Adaptec read benchmark, or specific write
benchmarks.

We would like to migrate to a more open solution, allowing us to
benchmark different disk types (the above only benchmark SCSI),
and to simulate some specific conditions. We already use
e.g. Bonnie through the filesystem, but this has the following
problems:

- higher latency (this could be fixed through the use of a low-overhead
filesystem, at least partly, but I haven't re-found it)
- caching (we don't know how exactly this gets into our way)

Using the /dev/sdX interface shows very bad performance (because one
request goes through user mode at a time, it seems). Using /dev/sgX
has limitations and is quite slow.

I thought I could try the following:
- call the request_fn() function of the device myself from a loadable
module (assuming this device is otherwise idle) with buffers allocated
by me.

Is there any counter-indication of doing this ? What are the precautions
to take ? Is there any more obvious way to do this ?

Thank you for any idea, pointers, and comments.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu