BM Seer Unofficial thoughts from an anonymous Sun employee

RAID Rebuild Performance Sun Storage 7110 Unified Storage Array

Monday Nov 10, 2008

The Sun Storage 7110 Unified Storage Array can rebuild RAID sets in less than an hour at 100% disk capacity, which is much faster than competitive offerings.

The Sun Storage 7110 Unified Storage Array minimizes rebuild time when recovering from replacing a drive and any other interruption than may cause RAID sets to rebuild. With the technology breaking use of ZFS to manage the file systems, rebuild times are some of the fastest in the industry.

Hard disk capacities are rapidly increasing and so are RAID rebuild times. Many storage administrators are using fewer and fewer drives in an array groups.

At 10% total capacity used the Sun Storage 7110 Unified Storage Array rebuilds RAID sets in only 6 minutes if your total capacity is at 10% used.

At 100% total capacity used the Sun Storage 7110 Unified Storage Array rebuilds RAID sets in 54 minutes.

The Sun Storage 7110 Unified Storage Array rebuilds 5x faster than EMC and NetApp.

The Sun Storage 7210 Unified Storage Array with SATA drives can rebuild more than 2x faster than EMC or NetApp.

For the test presented, the rebuild process was fairly linear with capacity used.

Competitive Landscape

Rebuild times for 500GB SATA, 80% full
Rebuild Time - Idle System No load
Storage Time Data Protection
EMC CX3-40 5 hrs 1 min RAID 5
NetApp FAS3050c 4 hrs 53 mins RAID 4
Pillar Axiom 500 3 hrs 48 mins RAID 5
SS7210 2 hrs 8 mins RAID Z
SS7110 (SAS, 146GB, 100% full) 54 mins Mirroring

Benchmark Description

In this test, a Sun Storage 7110 Unified Storage Array was populated with over 3 million files (each file was 150k). Then one of the drives was pulled and the rebuild was timed.

See Also

Disclosure Statement:

Sun Microsystem generated RAID rebuild results. Results reported 11/10/08.

Results Summary (Results are based on disk access performance with client caches enabled).
% Disk Capacity Used Protect Level Time (minutes)
10 Mirroring 6
100 Mirroring 54
5 RAIDZ 8

[14] Comments
Like this post? del.icio.us | furl | slashdot | technorati | digg
Posted at 08:33AM Nov 10, 2008 in Benchmark Issues
Tags:
Comments:

Sorry, but this test makes no sense IMHO.
It doesn't show the most important metric - how much IO performance is degraded during rebuild.
If IO completely stalls for 1 hour it is much worse than if response time is increased by 20% for 3 hours. This arrays are not your average portable hard drive so the business impact could be quite noticeable.

Posted by Mike on November 10, 2008 at 09:11 PM PST #

This is a scenario that can happen, so for some the test makes sense.

Yes it will be interesting to see the impact on IO during a rebuild. I heard that information will be coming soon. I'll of course post it if I see it.

Given the design of the Sun Storage 7000 family & ZFS I think that the response time will show a lot less impact than on competitive offerings. Hopefully data soon.

Posted by BM Seer on November 11, 2008 at 08:08 AM PST #

Thanks.
Very promising product.

Posted by Mike on November 11, 2008 at 06:30 PM PST #

I'm always pleased when someone references my work (Demartek) in an article or blog entry! I have a few comments on the results that you have posted.

I agree with the other commenter regarding the specific test posted, that of rebuild time on an idle system. Rarely in production environments is a storage system completely idle. This is why we performed six different sets of tests in our published results. We used the idle system test as a baseline, but then ran five other different load tests to show the important metric of how much I/O degradation occurs during a rebuild.

Secondly, the tests we performed were all against storage systems performing their RAID functions in hardware. As such, the results obtained from these hardware-based RAID storage systems are largely independent of operating systems used on the hosts driving the I/O workloads. As I understand it, RAID-Z avoids the traditional “write penalty” of standard RAID-5, but RAID-Z is a feature of ZFS that involves host server processor cycles, and so this is not a direct “apples-to-apples” comparison. When showing a test using ZFS, the processing power and application load on the host server must also be noted.

Third, the test showing mirrored SAS 15K RPM drives isn’t even close to a good comparison against systems using SATA drives using RAID-5 or technologies similar to RAID-5. Obviously, the 146GB SAS 15K RPM drives have significant performance advantages over SATA 7200 RPM drives. Further, rebuilding a mirrored drive is, essentially, a data copy with little or no intervention from the RAID controllers. Rebuilding a drive in a RAID 5 (or similar) disk group is far more complex and involves performing parity calculations, read/write/verifying the data, etc. The more drives in a parity-based disk group the longer the rebuild will be. Based on the data you show, one could extrapolate quite lengthy rebuild times to rebuild 500GB of data even on the SAS 15K RPM drives.

Your results also do not fully describe how the tests were conducted and based on what you have posted, cannot be independently verified.

I would be more than happy to work with Sun to do an independent 3rd-party comparison of your products drive rebuild times.

Posted by Dennis Martin on November 12, 2008 at 10:46 AM PST #

Dennis: I didn't run this particular test, I was reporting the work of other engineers. I've seen some internal chatter among various performance engineers including some who were involved. I don't think it will be too long before other tests will be made public to answer some of your questions.

I'm an engineer and don't deal with the sort of external test you propose, so I have no idea how those sorts of things are set up.

I did hear that The Sun Storage 7000 family is open to free try-and-buys. For more info on that program see... http://www.sun.com/tryandbuy/

Posted by BM Seer on November 12, 2008 at 02:54 PM PST #

Thanks. I'll look for those results to see how they compare.

Posted by Dennis Martin on November 12, 2008 at 03:01 PM PST #

This is one of the most ridiculous comparisons I've ever seen. Anybody even remotely familiar with storage sees what Sun did with this Apples-to-Oranges comparison. Sadly, I don't think this was done on purpose by Sun. Sun has proven time and time again that they don't understand storage. This is just another example. The post by Dennis Martin was dead-on:
15K SAS vs 7,00 SATA
146GB vs 500GB
RAID 1 vs RAID 5

What a joke.

Posted by Joe on November 12, 2008 at 05:44 PM PST #

This is just the start of the data coming. Sun Storage 7110 Unified Storage Array will have lots more data coming.

I know the people looking for data on NetApps & EMC had a hard time finding data - data from multiple sources on this topic are virtually non-existent.

This is just the beginning of performance data. I've already seen more internal data that didn't make it in time for launch. For what is public now take a look at: http://blogs.sun.com/bmseer/entry/perf_blogs_on_the_sun

Another point not to forget is this is an Open Solaris based appliance software: open source ZFS, DTrace, and Zones...

Least we forget lack of those excessive SW fees you see on other offerings.

Posted by BM Seer on November 13, 2008 at 06:43 AM PST #

"lack of those excessive SW fees"
You get what you pay for. Sun has never been good at storage. That's not to say they don't sell good storage (gear they OEM from LSI is good, basic storage) but their support is horrible for storage. And let's not forget their history: The T3, the DotHill stuff (3510/3511), the 6920. Disasters.

Sun should just focus on what it's good at and sell the LSI gear in the low-end and mid-range as part of their server deals. Sell the HDS in the high end. The stuff they make is junk.

Posted by Joe on November 13, 2008 at 11:21 AM PST #

Joe you make no sense. Who are you trying to protect? :)

The current NAS vendors now have some competition. Want to see how it works at your site there is a try & buy program.

past doesn't indicate the future, especially when the technology changes...

Apollo 1 disaster - NASA should stop trying to go to the moon.

Mars Observer was one of three NASA Mars missions lost in the 1990s, but look at those Rovers today! and the fine long service of Phoenix lander.

Posted by BM Seer on November 13, 2008 at 11:51 AM PST #

BM,
Sun is a server company. Period. If what they did to STK doesn't convince you of that then nothing else will.

Posted by Joe on November 13, 2008 at 01:21 PM PST #

Sun is many things my friend, a software company, a server company, and yes a storage company. Also storage is changing (maybe faster than you imagine). More announcements coming.

I think you need to explore the 7000 series on its merits, especially when compared to other NAS solutions.

Posted by BM Seer on November 13, 2008 at 02:04 PM PST #

Odds are it's using SAMBA.

Posted by Joe on November 13, 2008 at 04:04 PM PST #

It is not using Samba. Please refer to the in-kernel CIFS project at http://www.opensolaris.org/os/project/cifs-server/ or read this blog entry at http://blogs.sun.com/amw/entry/cifs_in_solaris or check out Ben Rockwood's blog entry (and he knows a thing or two about storage) at http://www.cuddletech.com/blog/pivot/entry.php?id=987 where he talks about using the FISHworks for almost 2 years.

Posted by Wayne Abbott on November 13, 2008 at 04:56 PM PST #

Post a Comment:
Comments are closed for this entry.