CSCAN vs CFQ I/O scheduler benchmark results

CSCAN vs CFQ I/O scheduler benchmark results

[Vishal Patil]

Hello

I ran the sysbench benchmark to compare the CSCAN I/O scheduler
against the CFQ scheduler and following are the results. The results
are interesting especially in case of sequential writes and the random
workloads

                                            Latency (seconds)

               seq            seq         seq             rnd
 rnd         rnd
               reads         writes      r + w          reads
writes      r + w
               --------------------------------------------------------------------------------------
CFQ         0.0116      0.0164      0.0107      0.1178       0.0423      0.0605

CSCAN    0.0148      0.0092      0.0169      0.1043      0.0473      0.0732


                                           Throughput (MB/seconds)

                seq            seq         seq             rnd
  rnd         rnd
                reads         writes      r + w          reads
writes      r + w

--------------------------------------------------------------------------------------
CFQ        19.062      15.251      22.127      2.1197      1.0032       1.376

CSCAN   14.553      22.108      14.72       2.394       0.9304         1.399


The machine configuation is as follows
CPU: Intel(R) Pentium(R) 4 CPU 2.80GHz
Memory: 1027500 KB (1 GB)
Filesystem: ext3
Kernel:   2.6.16.2

If interseted you may have a look at the raw data at
http://www.google.com/notebook/public/14554179817860061151/BDQtXSwoQ2_mdxLgh

- Vishal

CSCAN vs CFQ I/O scheduler benchmark results

[Vishal Patil]

The previous mail got scrambbled and hence I am resending this one
again

Hello

I ran the sysbench benchmark to compare the CSCAN I/O scheduler
against the CFQ scheduler and following are the results. The results
are interesting especially in case of sequential writes and the random
workloads

                                            Latency (seconds)

               seq            seq         seq             rnd
  rnd          rnd
               reads         writes      r + w          reads
writes      r + w
               --------------------------------------------------------------------------------------
CFQ         0.0116      0.0164      0.0107      0.1178       0.0423      0.0605

CSCAN    0.0148      0.0092      0.0169      0.1043      0.0473      0.0732


                                           Throughput (MB/seconds)

                seq            seq         seq             rnd
 rnd         rnd
                reads         writes      r + w          reads
writes      r + w

--------------------------------------------------------------------------------------
CFQ        19.062      15.251      22.127      2.1197      1.0032       1.376

CSCAN   14.553      22.108      14.72       2.394       0.9304         1.399


The machine configuation is as follows
CPU: Intel(R) Pentium(R) 4 CPU 2.80GHz
Memory: 1027500 KB (1 GB)
Filesystem: ext3
Kernel:   2.6.16.2

If interseted you may have a look at the raw data at
http://www.google.com/notebook/public/14554179817860061151/BDQtXSwoQ2_mdxLgh

- Vishal


-- 
Success is mainly about failing a lot.

Re: CSCAN vs CFQ I/O scheduler benchmark results

[Jan Engelhardt]

> The previous mail got scrambbled and hence I am resending this one
> again
>
Still. Here are the values, fit for 80 cols.

			Latency (seconds)
			CFQ	CSCAN
	seq reads	0.0116	0.0148
	seq writes	0.0164	0.0092
	seq r+w		0.0107	0.0169
	rnd reads	0.1178	0.1043
	rnd writes	0.0423	0.0473
	rnd r+w		0.0605	0.0732

			Throughput MB/sec
			CFQ	CSCAN
	seq reads	19.062	14.553
	seq writes	15.251	22.108
	seq r+w		22.127	14.72
	rnd reads	2.1197	2.394
	rnd writes	1.0032	0.9304
	rnd r+w		1.376	1.399

Excels at sequential writes, so it seems like a good idea to use CSCAN on a 
target drive when doing `rsync disk1/ disk2/`, esp. for large 
files.



Jan Engelhardt
-- 

Re: CSCAN vs CFQ I/O scheduler benchmark results

[Jens Axboe]

On Fri, Jun 09 2006, Vishal Patil wrote:
> The machine configuation is as follows
> CPU: Intel(R) Pentium(R) 4 CPU 2.80GHz
> Memory: 1027500 KB (1 GB)
> Filesystem: ext3
> Kernel:   2.6.16.2

You don't mention the storage used, which is quite relevant.

If you have the time, please rerun with 2.6.17-rc6-gitX latest. Although
I'm not sure why you think CSCAN is a good scheduling algorithm, in
general it may be fine but there are trivial non-root 'dos' attacks. Any
of the non-noop Linux io schedulers is a better choice imo.

-- 
Jens Axboe

Re: CSCAN vs CFQ I/O scheduler benchmark results

[Vishal Patil]

Jan

I ran the performance benchmark on an IBM machine with the following
harddrive attached to it.

cat /proc/ide/hda/model
ST340014A

Also note the CSCAN implementation is using rbtrees due which the time
complexity of the different operations is O(log(n)) and not O(n) and
that might be the reason that we are getting good values for specially
in case of sequential writes and the random workloads.

I will try making measurements using 2.6.17-rc6-gitX until next weekend.
Thanks for help and inputs folks.

- Vishal

On 6/11/06, Jens Axboe <axboe@suse.de> wrote:
> On Fri, Jun 09 2006, Vishal Patil wrote:
> > The machine configuation is as follows
> > CPU: Intel(R) Pentium(R) 4 CPU 2.80GHz
> > Memory: 1027500 KB (1 GB)
> > Filesystem: ext3
> > Kernel:   2.6.16.2
>
> You don't mention the storage used, which is quite relevant.
>
> If you have the time, please rerun with 2.6.17-rc6-gitX latest. Although
> I'm not sure why you think CSCAN is a good scheduling algorithm, in
> general it may be fine but there are trivial non-root 'dos' attacks. Any
> of the non-noop Linux io schedulers is a better choice imo.
>
> --
> Jens Axboe
>
>


-- 
Success is mainly about failing a lot.

Re: CSCAN vs CFQ I/O scheduler benchmark results

[Jens Axboe]

(please don't top post)

On Sun, Jun 11 2006, Vishal Patil wrote:
> Jan
> 
> I ran the performance benchmark on an IBM machine with the following
> harddrive attached to it.
> 
> cat /proc/ide/hda/model
> ST340014A

Ok, so plain IDE.

> Also note the CSCAN implementation is using rbtrees due which the time
> complexity of the different operations is O(log(n)) and not O(n) and
> that might be the reason that we are getting good values for specially
> in case of sequential writes and the random workloads.

Extremely unlikely. The sort overhead is completely noise in a test such
as yours, an O(n^2) would likely run just as fast.

-- 
Jens Axboe

Re: CSCAN vs CFQ I/O scheduler benchmark results

[Vishal Patil]

Jens

Could you let me know what tests would be fair to make comparsion
between the I/O schedulers? Thanks.

- Vishal

Re: CSCAN vs CFQ I/O scheduler benchmark results

[Nate Diller]

On 6/12/06, Vishal Patil <vishpat@gmail.com> wrote:
> Jens
>
> Could you let me know what tests would be fair to make comparsion
> between the I/O schedulers? Thanks.

any filesystem benchmark should be informative, namesys tests with a
whole suite, bonnie++, iozone, mongo (custom), and one or two others.
You can ask on reiserfs-list for more information.

I suggest testing against more than one filesystem as well, it's quite
common for a FS to "prefer" one scheduler.

NATE

Re: CSCAN vs CFQ I/O scheduler benchmark results

[Jens Axboe]

On Mon, Jun 12 2006, Vishal Patil wrote:
> Jens
> 
> Could you let me know what tests would be fair to make comparsion
> between the I/O schedulers? Thanks.

Depends on what you want to test, of course! I can't give you an answer
on that.

The thing about IO schedulers is that it's easy to provide good
throughput or good latency, but hard to do both. The tests you did so
far have x processes doing the same thing. Try something that has eg an
async writer going full throttle, and then a/some sync readers.

-- 
Jens Axboe