UNSOLVED

JVBowling

updated

16 years ago

J

JVBowling

1 Rookie

•

32 Posts

0

2683

July 13th, 2010 06:00

TDEV Performance

We have a new VMAX and have started testing performance of our system (running a few benchmark tests prior to moving into production)...

My understanding is an extent (768KB) is issued against a given Data Device

A Data Device is created against the underling RAID Disk Group (i.e., RAID 5 (3+1))

So, if I happen to have a pool of 252 drives configured in a RAID 5 (3+1) configuration, and require a 140GB Data Volume, would it be best to:

#1. Just create a large TDEV and be done with it...?

or

#2. Create a 144GB Meta from (3 x 48GB TDEV devices)?

48GB TDEV came from 252 drives \ 4 = 63 data devices that fit across the RAID Groups

                                                          63 x 768GB Extent Sizes = 48GB

Am I over complicating this or is this a good practice for long-term growth?

Just wanted to run this by the group before I start the next set of tests...

Thanks!

  • Quincy561

    1281 Posts

    1213

    1

    Posted July 13th, 2010 07:00

    There is an excellent presentation on VP performance considerations that was given at EMC World this year.  A copy of it is posted to the Symmetrix SPEED web site.  Please talk you your local EMC account team about getting a copy.

  • JVBowling

    1 Rookie

    •

    32 Posts

    1212

    0

    Posted July 13th, 2010 08:00

    I've sent them the request for this paper.

    Thanks for the response.

  • Quincy561

    1281 Posts

    1212

    1

    Posted July 13th, 2010 10:00

    That was the one.

    For TDATs, you need enough drives to support the workload.  It really isn't much different than sizing a Symm for performance with thick provisioning.  RAID choice will effect

    For the TDEV, you may find that more TDEVs will improve performance, and because of this striped meta volumes can provide greater performance than concatenated meta volumes.

  • JVBowling

    1 Rookie

    •

    32 Posts

    1212

    0

    Posted July 13th, 2010 10:00

    I was given "Virtual Provisioning Best Practices for Performance on Symmetrix VMAX" by Tom Rogers.

    This is an excellent document, but it doesn't address the specific's regarding optimal TDEV sizing based upon the number of Data Devices and Drives per RAID Group.  Maybe there's another document that's out there?

    I should have initial results with large TDEVs vs. Meta TDEVs to see if there's a performance difference we see with SQLIO and IOMeter.

    Thanks for everyone's feedback

  • Booyah2

    2 Intern

    •

    185 Posts

    1212

    1

    Posted July 13th, 2010 11:00

    Here's another VP Best Practice Implementation Guide...find attached.

    This may help with your implementation.

    1 Attachment

  • JVBowling

    1 Rookie

    •

    32 Posts

    1212

    0

    Posted July 16th, 2010 05:00

    I was "expecting" tests to conclude Concantenated Metas were as good, if not better, than Striped Meta's in the "Virtual TDEV\TDAT World"...

    So far, you are correct.  I see significant IOPS and MB's improvement with TDEV's configured exactly the same way (other than Striped Meta's vs. Concatenated). Using both SQLIO and IOMeter.

    I also see slightly better performance (but nothing overwhelming) using more\smaller TDEV's to form Meta's...

    My assumption in the original question was I would get better perfomace if I created TDEV's that fit nicely with the given number of drives instead of large ones that wrap around and could potentially access the same set of drives more than once.

    So far, I've been testing smaller files (5MB - 50GB)...  Today, I'm running tests against 100GB files on 800GB Meta's to see if  the results are similar.

    Nice to have a couple weeks to test before going into production (this doesn't happen very often)...

    So far, I've concluded Striped Meta's are better for those we likely will not expand in the future: 800GB ESX LUNs come to mind.

    Having 20 x 40GB TDEVs vs 4 x 200GB TDEVs only provided trivial performance improvements so keeping it simple may be best.

    If further testing proves different, I'll post.

    I'd also be interested in hearing what others have experienced.