PowerVault 220s & Windows 2003 Cluster Performance

Question

:smileysad: We are having big problems with our system. A PowerEdge 2850 & PV220s (clustered) - Using Perc4DC.   We have a lot of these servers (in normal configuration, i.e. not clustered) and haven't had any major issues.  We have setup a new clustered system (all patched and firmwared) and are having massive performance issues.  Everytime someone copies any data of any significant size the system stops dead. (also all client computers with drives mapped to it slow or stop).   I have read lots of posts about performance being slower when clustering but the kind of speeds we are getting seems to be a bit of a joke.   We have run NBench on a range of systems and the results are very worrying.   On a system not clustered we are getting these results (PE2850 + PV220s + Perc4DC)   (With Raid 5 write back cache turned on) Write - 65 MBytes/sec Read - 110 MBytes/sec   (With Raid 10 write back cache turned on) Write - 63 MBytes/sec Read - 114 MBytes/sec   On a system Clustered we are getting these results (The same hardware as above)   (With Raid 5 write through cache turned on) Write - 6 MBytes/sec Read - 106 MBytes/sec   (With Raid 10 write through cache turned on) Write - 7.9 MBytes/sec Read - 114 MBytes/sec .......   Now correct me if I am wrong, but isn't the performance difference very poor.  After reading some of the posts, I expected the speed to drop a little bit but not 10 times slower?   My questions are.... What can be done to resolve this? Will it mean buying new hardware? Is there a way of getting much better speeds with a config change on the current hardware? Where does the actual problem lie? (In the PE2850 or the Perc4DC or the PV220s?).   What we are looking for is Highly available and not Hardly available!   Thanks to anyone who can help.

660man · Answer

Tim

what is your config? ie number of disks ? size of disks? and number of disks in your LD

I have performed extensive testing on perc's and write performance when clustered is dramatically slower, because basically when clustered you can not be performing write cacheing on a scsi solution where the controller is in the host..

the figures you are getting look about right based on my testing i have done previously

however... if you look at FC systems you can have write cache when clustered because the cache is on the array

660man · Answer

your issue lies in the fact that when you enable clustering, your cache policy is changed for you to write through not write back.   write back cahce is basically cache enabled whereas write through is basically no write cache.   this is forced in cluster mode so that you dont get data corruption in the event of a failover.   Check out the readme files on the latest firmware for your perc.. contains updates for clustered performance issues.

Tim Arnold · Answer

The Perc4DC controller has been updated to the newest firmware (351S)

I have had a look at the readme file. and all it mentions is that it has fixed some problems.

What is the suggested config for a cluster? Also what would you say would be an "acceptable" write speed?

Tim Arnold · Answer

PE2850 PV220s Perc4DC 9 x 146GB 10k Segate (have had it at raid 5 and raid 10) 2 x 73GB 10k Maxtor (quorum)   what do you think

660man · Answer

HI tim

and your driver is 6.46.2.3?

looks fine and performance is roughly the same as i was seeing..

I know it is a huge jump

You can re test if you like.. set the perc to non cluster and set the cache to write through.( no cahce)

you will get exactly the same results.

As i said it is a limitation of MSCS with SCSI clusters that you have no write cache.

This limitation is across all vendors not just dell.

It is not there when using Fibre Arrays where the Raid Inteligence is on the array not in the Host.

If you need the write speed then SCSI clusters are not what you need you may want to look at Fibre arrays such as Dell|EMC Clarrion CX range

Tim Arnold · Answer

Doh!

Looks like were stuffed then.

It was nice of the Dell sales team to warn us of the performance difference then.

We are going to be moving to a SAN but not for a while.

I think our new plan of action is going to be to turn the cluster off and just run one server until the SAN arrives.

Thanks for everyones help

jasonkappy · Answer

You guys are lucky getting atleast 2 digit reads. I get a max of 7 Mbytes/sec reads. We have Dell Gold support. That guy checked all the parameters and said they were ready to replace cables,ZIMM ,Controller.....what not? Blind dart shot. That guy told me we should expect atleast 30MBytes/sec writes and 50 MBytes/sec. This was official tests done by Dell with Windows 2000 and SQL server.Did not play around changing components as the machine was due production in 2 days. Machine has 120GB oracle hotbackup dumps and takes around 4 hrs to take that backup(netbackup)..big pain.....Did any one try to upgrade to new version 351S? Any difference? What is your cache read policy?   Thanks, J.

Tim Arnold · Answer

Hmmmm. Sounds like you are having the same Dell issues as we are.

The guy from Dell that told you that you should get at least 30mb/sec write and 50mb/sec read... Did you get that in wrinting?/Email? Wouldnt mind getting his email address.... so we can find out the magic switch that speeds everything up.

We are running on the newest firmware (351S) and to be honest I think it slowed performance down even more.

We have had the controller set to all of the options... the cache read policy the last time we tried was set to..i think it is set to write through...

hope it helps

jasonkappy · Answer

@Tim Arnold wrote:

Hmmmm. Sounds like you are having the same Dell issues as we are.

The guy from Dell that told you that you should get at least 30mb/sec write and 50mb/sec read... Did you get that in wrinting?/Email? Wouldnt mind getting his email address.... so we can find out the magic switch that speeds everything up.

We are running on the newest firmware (351S) and to be honest I think it slowed performance down even more.

We have had the controller set to all of the options... the cache read policy the last time we tried was set to..i think it is set to write through...

hope it helps

He said that he can only read those statistics. He cannot mail those stats as they are only intended for Dell internal use. He will be fired if he had given them to me in written/email copy.

Thanks,
J.

dlippi · Answer

This is *not* a problem with all vendors.  I know of a major vendor that has battery-backed cache from controllers that are not installed in the hosts.  Not sure if I can put in another vendor's Hearty Products or not on this forum.

660man · Answer

Dell also have solutions where the cache is battery backed in an eternal array.. thats my point.   with scsi clusters the controller and cache and batter backup is inside the HOST server.   so therefore the cache is set to write through (aka no write cache) when clustered... which you will find is the same accross the board of vendors.

PowerVault

PowerVault 220s & Windows 2003 Cluster Performance

Was this post helpful?