Andrea Arcangeli <firstname.lastname@example.org> writes:
> On 22 May 2000, Zlatko Calusic wrote:
> >Question for Andrea: is it possible to get back to the old speeds with
> >tha new elevator code, or is the speed drop unfortunate effect of the
> >"non-starvation" logic, and thus can't be cured?
> If you don't mind about I/O scheduling latencies then just use elvtune and
> set read/write latency to a big number (for example 1000000) and set the
> write-bomb logic value to 128. However in misc usage you care about
> responsiveness as well as latency so you probably don't want to disable
> the I/O scheduler completly. The write bomb logic defaul value is too
> strict probably and we may want to enlarge it to 32 or 64 to allow SCSI
> to be more effective.
Yes, you are right. I tested performance with different parameters to
elvtune, and it definitely improves rewriting speed if I put bigger
numbers. In fact with 1000000/1000000/128 it is back to 2.3.42 results
(rewriting speed ~9.2MB/s).
While at it, could you explain what are write bomb logic and latency
numbers? What do they define?
> About the bad VM performance of the latest kernels please try again with
> pre9-1 + classzone-28.
I did a test with pre8 + classzone28 (didn't have 9-1 at the moment,
but there was only one easily resolvable conflict in vmscan.c). And I
must admit that I'm completely satisfied with its behaviour. System is
definitely working right AND fast. It reads, writes and swaps with
full speed and the system survived all tests I made against it. With
classzone patch applied, VM/IO balance/behaviour seems perfect.
Too bad the patch is not going to make it to the Linus kernel tree,
but hey, there's always a possibility of applying a patch for better
performance. Assuming you continue porting the patch to forthcoming
kernels, of course (that's to say I tried to port it to the pre9-4 and
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to email@example.com Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Wed May 31 2000 - 21:00:11 EST