RE: [RFC] add sysfs to dynamically control blk request tag maintenance

["Chen, Kenneth W" <kenneth.w.chen@intel.com>]

Jens Axboe wrote on Friday, October 07, 2005 1:08 AM
> > It's probably a very small number that I'm chasing with avoiding blk
> > layer tagging.  Nevertheless, any number no matter how small, is a
gold
> > mine to me :-)
> > 
> > Latest execution profile taken with 2.6.14-rc2 kernel with "industry
> > standard transaction processing database workload".  First column is
> > clock ticks (a direct measure of time), 2nd column is instruction
> > retired,
> > and 3rd column is number of L3 misses occurred inside the function.
> > 
> > Symbol			Clockticks	Inst. Retired	L3
Misses
> > scsi_request_fn		8.12%	9.27%	11.18%
> > Schedule			6.52%	4.93%	7.26%
> > scsi_end_request		4.44%	3.59%	6.76%
> > __blockdev_direct_IO	4.28%	4.38%	3.98%
> > __make_request		3.59%	4.16%	3.47%
> > __wake_up			2.46%	1.56%	3.33%
> > dio_bio_end_io		2.14%	1.67%	3.18%
> > aio_complete		2.05%	1.27%	3.56%
> > kmem_cache_free		1.95%	1.70%	0.71%
> > kmem_cache_alloc		1.45%	1.84%	0.45%
> > put_page			1.42%	0.60%	1.27%
> > follow_hugetlb_page	1.41%	0.75%	1.27%
> > __generic_file_aio_read	1.37%	0.36%	1.68%
> 
> The above looks pretty much as expected. What change in profile did
you
> see when eliminating the blk_queue_end_tag() call?

I haven't make any measurement yet.  The original patch was an RFC, and
I want to hear opinions from the experts first.  I will do a measurement
with the change in qla2x00 driver and I will let you know how much
difference does it make.  Most likely, clock ticks in scsi_request_fn
and
scsi_end_request should be reduced.  It will be interesting to see L3
misses stats as well.

- Ken
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/