zcache preliminary benchmark results

zcache preliminary benchmark results

[Dan Magenheimer]

Last November, in an LKML thread I would rather forget*, James
Bottomley and others asked for some benchmarking to be done for
zcache (among other things).  For various reasons, that benchmarking
is just now getting underway and more will be done, but it might be
useful to publish some interesting preliminary results now.

Summary: On a kernel compile "make -jN" workload, with different
values of N to test varying memory pressure, zcache
shows no performance loss when memory pressure is low,
and up to 31% performance improvement when memory pressure
is moderate to high.  RAMster does even better.

(Note that RAM is intentionally constrained to 1GB to force
memory pressure for higher N in the workload.)

* thread summarized in LWN (http://lwn.net/Articles/465317/)

=========

Benchmark results and description:

(all results in seconds so smaller is better)
N=	nozcache	zcache	faster by	RAMster	faster by
4	879		877		0%		887		-1%
8	858		856		0%		866		-1%
12	858		856		0%		875		-2%
16	1009		922		9%		949		6%
20	1316		1154		14%		1162		13%
24	2164		1714		26%		1788		21%
28	3293		2500		31%		2177		51%
32	4286		4282		0%		3599		19%
36	6516		6602		-1%		5394		22%
40	DNC		13755				8172		68% (over zcache)

DNC=did not complete: stopped after 5 hours = 18000

Workload:
	kernel compile "make -jN" with varying N
	measurements in elapsed seconds
	boot kernel: 3.2 + frontswap/ramster commits
	Oracle Linux 6 distro with ext4
	fresh reboot for each test run
	all tests run as root in multi-user mode

Hardware:
	Dell Optiplex 790 = ~$500 (two used for RAMster)
	Intel Core i5-2400 @ 3.10 GHz, 4coreX2thread, 6M cache
	1GB RAM DDR3 1333Mhz (for RAMster, other server has 8GB)
	One 7200rpm SATA 6.0Gb/s drive with 8MB cache
	10GB swap partition

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: zcache preliminary benchmark results

[Seth Jennings]

On 03/21/2012 06:30 PM, Dan Magenheimer wrote:
> Last November, in an LKML thread I would rather forget*, James
> Bottomley and others asked for some benchmarking to be done for
> zcache (among other things).  For various reasons, that benchmarking
> is just now getting underway and more will be done, but it might be
> useful to publish some interesting preliminary results now.

I'd also like to post some zcache performance numbers that suggest
that zcache makes an even more impressive change to the amount
of total I/O when the system is under light to moderate memory
pressure.

Test machine:
Gentoo w/ kernel v3.3 + frontswap (cleancache disabled)
Quad-core i5-2500 @ 3.3GHz
1GB DDR3 1600MHz (limited with mem=1024m on boot)
Filesystem and swap on 2x80G RAID0
majflt are major page faults reported by "time"
pswpin/out is the delta of pswpin/out from /proc/vmstat before and after
the make -jN

Each run started with with swapoff/on and
echo 3 > /proc/sys/vm/drop_caches

I/O									
	normal				zcache				change
N	pswpin	pswpout	majflt	I/O sum	pswpin	pswpout	majflt	I/O sum	%I/O
8	0	133	1781	1914	0	0	1835	1835	4%
12	10	1140	1871	3021	0	5	1886	1891	37%
16	675	1978	2330	4983	21	63	3771	3855	23%
20	3420	6197	3421	13038	265	786	5218	6269	52%
24	28358	51884	8865	89107	3944	6227	36048	46219	48%
28	44132	62182	11931	118245	6094	11362	74323	91779	22%
32	94163	125086	22526	241775	22534	32803	179164	234501	3%
									
Runtime									
N	normal	zcache	%change						
8	284	280	1%						
12	283	281	1%						
16	281	280	0%						
20	289	310	-7%						
24	322	311	3%						
28	347	325	6%						
32	437	378	14%						
									
%CPU utilization (out of 400% on 4 cpus)
N	normal	zcache	%change						
8	245	245	0%						
12	249	251	-1%						
16	252	252	0%						
20	247	255	-3%						
24	221	230	-4%						
28	204	219	-7%						
32	162	187	-15%						

Some of my runtime curve N vs %change doesn't match up with
Dan's probably due to differing swap device speeds and I think
my .config had less to build so the runtime magnitudes are less.

Runtime %change will be effected by the swap device speed. But
the I/O reductions are less related to swap device speed and, IMHO,
really show the value of zcache.

Environments with shared storage could particularly like this.
You could enable swap on your machines and frontswap+zcache
would give you an early warning system for memory pressure.
If frontswap starts picking up pages, the admin can get a
warning and zcache will mitigate the swap I/O impacting the
SAN while something is done to relieve the memory pressure.

--
Seth

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>