Subject: New device buffer queue strategy
To: None <>
From: Juergen Hannken-Illjes <>
List: current-users
Date: 09/02/2002 16:05:40
A new device buffer queue strategy is available for testing.

The current implementation of disksort (a min seek sort) doesn't perform
very well with long queues of outstanding requests. Especially with
softdeps this is a very common situation. As all requests are ordered the
read latency may become very high.

The new strategy keeps disksort for write-requests while read-requests are
put on a second queue with first-come first-serve strategy.
Read-requests are served before write-requests.

While the overall I/O performance is the same, the system becomes more
responsive under high I/O load.

To test this strategy:

- lookup the `bufq_alloc' call in your favorite disk driver:

  ! sys/dev/scsipi/sd.c: bufq_alloc(&sd->buf_queue, BUFQ_DISKSORT|BUFQ_SORT_RAWBLOCK);

- replace the BUFQ_DISKSORT argument by BUFQ_READ_PRIO.

  ! sys/dev/scsipi/sd.c: bufq_alloc(&sd->buf_queue, BUFQ_READ_PRIO|BUFQ_SORT_RAWBLOCK);

- recompile kernel and test.

Comments and improvements are welcome.
Juergen Hannken-Illjes - - TU Braunschweig (Germany)