| Title: | HSJ30/40 Product Conference |
| Moderator: | SSDEVO::EDMONDS |
| Created: | Tue Jul 13 1993 |
| Last Modified: | Fri Jun 06 1997 |
| Last Successful Update: | Fri Jun 06 1997 |
| Number of topics: | 1264 |
| Total number of notes: | 4958 |
a large batch application is heavily IO bound, (6% cpu and 70 8KB read
IOS/sec to a few RDB tables on a stripeset)
as the table access is sequential, is it possible to tune the read
ahead on the HSJXX so that each IO is done in cache instead of disk ?
each of the 70 reads/sec takes 12 msec (avg) according to PSPA, and
that's the reason why the job takes so long..
70 times 12 = 840msec and the rest is CPU.
q1) how do i tune Read ahead for the HSJ?
Q2) shouldn't the *disk* cache buffer work in this case?
Bruno
| T.R | Title | User | Personal Name | Date | Lines |
|---|---|---|---|---|---|
| 1254.1 | What cache? | EYLAK::BATES | Ken Bates | Tue May 20 1997 11:15 | 21 |
> q1) how do i tune Read ahead for the HSJ?
There is no read-ahead cache at the controller level on the HSJ. You might
consider disabling cache for the drive, since cache processing does add some
overhead, as well as purging data from cache for other drives, and it's
probably unlikely that the sequentially accessed data will be accessed again.
> Q2) shouldn't the *disk* cache buffer work in this case?
It should, but that depends on how effective the drive cache is, and whether
the I/O is synchronous or asynchronous. Disk read-ahead caching is only
effective if there are pauses in the I/O stream (e.g., synchronous I/O), since
asynch I/O implies continuous work to the disk, and the disk should transfer
data at it's native rate. We've measured sequential access to disk drives, and
you can get high data rates *if* you keep the drive 100% busy *or* you have a
disk that has a "good" caching algorithm.
Karen (HUNTRS::) Ripley has been running tests on sequential access streams to
disks, and should have lots more information on this.
- Ken
| |||||
| 1254.2 | synch, slow data rate | ROM01::OLD_CIPOLLA | Bruno Cipolla | Tue May 20 1997 16:28 | 7 |
re: .-1 reads appear to be synch.
the total data rate is around 500KB/sec (70 16 blocks IOs at 10-12msec
each)
EMC is touting automatic read ahead for their controllers
Bruno
| |||||