0

We have a Blade with EVA storage array with 3 disk groups. One disk group consists of FATA 7200 1TB 18 hard drives, RAID 5, total IOPs 1000-1300. And I get statistics of used IOPs for this disk group of higher than 2000 IOPs. How can that be? If this is real and true how does 1200 IOPs work?

Andrew Schulman
  • 8,811
  • 21
  • 32
  • 47
Javid
  • 13
  • 2

2 Answers2

2

As Shane mentioned sequential IO is generally faster than random IO but there's also one other thing in play here, caching. It's actually quite hard these days to benchmark a system in a way that truly gives you a baseline as there's caching all over the place; on the disks, the channel controllers, the SAN controllers, the HBAs, the OS etc. It's great that you care about this, many don't, but some of the unexpected performance gains will just be regular caching.

I do have some other thoughts however if that's ok. I work for just about the largest EVA/P6k buyer HP have, we have hundreds and thus I know my way around them, I know what I love and I know what's caused us problems along the way. The single biggest issue we've had with storage in the last decade and across a 180,000 employee business was those 1TB FATA disks. They're not rated for 24/7 usage, only 30% of the day and if you do run them for longer then they die very quickly. Your use case may well be fine with this, maybe only using those disks for a few hours a day but there's a big kicker - levelling. When one of those disks dies, and it will, you replace it and that starts a levelling process, that may run for more than 30% of the day, often causing more disk failure. This is what we saw, a single disk died and caused an avalanche of failures. We only ran with those disks for a year or less but saw about 60% of them fail during that time. HP were massively apologetic but did point out in the fine print that they were only rated for a 30% duty cycle and so we ended up just swapping them out, hundreds of them, for 'full duty cycle' disks - i.e. not FATAs. Since then we've obviously avoided them and have in fact now retired all of our non-P6k EVAs. I cannot stress how much I dislike that disk and would urge you to move from them ASAP.

Oh and on a smaller note those older EVAs perform much better if your disk groups work in blocks of 8 disks; i.e. 8, 16, 24, 32 etc. disks - rather than 18, it's just how they do the striping.

Hope this helps.

Chopper3
  • 101,299
  • 9
  • 108
  • 239
0

Do you mean that your storage is rated at 1000-1300 IOPS, but you're observing 2000 IOPS?

There doesn't seem to be anything out of the ordinary there, as small/seqential IOPS are much faster than random large operations. If that 1000-1300 rating is from the vendor, then they're probably targeting more of a worst-case random IO workload for that rating - observing more operations than that is normal.

Shane Madden
  • 114,520
  • 13
  • 181
  • 251
  • Yes I calculate IOPs count and gived me 1200 IOPs for Disk Group.And i am using Vmware Operation Manager.I saw Our virtual machines that located in that disk group uses more than 2000 IOPs even 3000.But how this works?Does it drop other processes?And I want to know does it influence to Virtual machines performance. – Javid Mar 07 '14 at 06:09
  • @Javid How did you calculate that 1200 number? – Shane Madden Mar 07 '14 at 06:16
  • it-memories.info/?page_id=129 18 FATA 1TB single RAID 5 – Javid Mar 07 '14 at 06:27
  • @Javid That calculator doesn't make it clear what kind of IOPS those would be - but like I said, the estimation is likely for a less ideal workload, with random operations which are slower and different write sizes. Some workloads will give you much higher IOPS numbers on those disks - it's not something to worry about. – Shane Madden Mar 07 '14 at 06:39
  • Thank you very much.One thing.I have a SQL Server database.and its performance is low.Does it can happen from these IOPs? – Javid Mar 07 '14 at 06:58