On Thu 28-04-11 19:46:04, Jeff Mahoney wrote:
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
On 04/28/2011 06:17 PM, Jan Kara wrote:
On Thu 28-04-11 13:35:13, Jeff Mahoney wrote:
-----BEGIN PGP SIGNED MESSAGE----- Jan Kara: - - patches.suse/readahead-request-tunables.patch This is dependent on SUSE-specific CONFIG_KERNEL_DESKTOP. General update of the defaults probably wouldn't pass. So I think we have to carry this one... Ok, I had this one in mind if Suresh has any luck getting the tunable defaults upstream - or if we end up working with sysctls instead. An interesting side-idea might be to add the ability for the sysctl utility to load a kernel-specific sysctl.conf that we ship with each flavor and then apply the system values on top of those, overriding the ones that collide. We'd get the benefits without the kernel patches. Yeah, it would be easiest if we could do such parameter tuning without a kernel patch.
[ Up for discussion ] - - patches.suse/file-capabilities-disable-by-default.diff [never submitted] - - patches.suse/connector-read-mostly - - patches.suse/ext3-barrier-default - - patches.suse/reiserfs-barrier-default Although I believe the above two would be sane defaults, I kind of feel it might be too late to change the defaults upstream. It would probably generate lots of bug reports of the type "my IO is suddently slower"...
I'm not convinced they actually *are* sane defaults for commodity hardware. Having to flush the entire cache for a barrier is a pretty heavyweight operation, especially with the sizes of caches growing on individual disks. In my own testing it seems like it's faster to disable the write cache and use regular buffer waits rather than flush the cache, but it's been a while since I've done that. I guess it depends on the drive and the load. I remember Tejun did some testing of this when he was working on the barrier code and found out that write caches are useful on normal SATA drives. The drive can still do more reordering of IO when it has write caches enabled. But I agree that e.g. for fsync heavy workloads, turning off write caches may well be the faster option.
Honza -- Jan Kara <jack@suse.cz> SUSE Labs, CR -- To unsubscribe, e-mail: opensuse-kernel+unsubscribe@opensuse.org For additional commands, e-mail: opensuse-kernel+help@opensuse.org