safe IOPS

 
SOLVED
Go to solution
Marcus Longardt
Advisor

safe IOPS

Hello,

what understand hp under safeIOPS?

And how is the calculation?

Regards
Marcus
17 REPLIES 17
Víctor Cespón
Honored Contributor

Re: safe IOPS

It's the amount of operations a group of disks can do without having a high latency. The limits are 15 ms for a read and 5 ms for a write.

This depends on the number of disks, what kind of drives they are and the kind of RAID (0/1/5/6).

For EVAs you have a table here:

http://h20000.www2.hp.com/bizsupport/TechSupport/Document.jsp?objectID=c01671044

Take the number for the type of disks and the RAID level you use, multiply by tne number of disks on the disk group.

The latency grows fast with the number of writes. Having 99% writes is much worse than having 99% reads, specially in RAID 5.
Marcus Longardt
Advisor

Re: safe IOPS

Hello vcespon,

thank you for your answers.

First I understand the hole message, that safeIOPS are depends from the latency (max 15ms write and 5ms read) and the max IOPS the diskgroup are able to realize against the used IOPS.

But how was that calculated?

The Link shows me the IOPS for the Disks - but I can't see how many percent safIOPS I have?

Mmh - perhaps you can concretize your answer?

Regards
Marcus
S. Boetticher
Regular Advisor

Re: safe IOPS

- the calculation for the given values seems to incorporate RAID-penalties (4 IOPS for 1 write in VR5, ...)
- you should multiply your number of HDDs in your diskgroup by the IOPS per HDD (choose the disk type and VRAID setting for your EVA), then you "know" the max IOPS for your system before latency will skyrocket and users starts to complain
- how to get "your" real IOPS:
* use evaperf to monitor, what happens with your EVA, how many IOPS hits which DG and which Controller and and and...
* use "calculators" for well-known apps like Exchange to estimate, how many IOPS they would put onto your array
S. Boetticher
Regular Advisor

Re: safe IOPS

@vcespon:

do you have a more detailed table for 100% random read I/O?

background: we don't have that much write IO, according to EVAperf it's easily being taken care of by the mirrored cache, i.e. we don't need to care for write penalty (VR1=*2, VR5=*4) in 99% of the day

but we face random read I/O:
I see worstcase 2500 random reads by exchange 2003 server on a 48x300GB10K DG consisting only of VR5 (logs are separate 8x300x10K, only a few IOPS), that are not coming from cache, i.e. is simply a matter of spindle count.

I read somewhere that a 10K drive could do approx. nearly 100 IOPS (without RAID penalty!) (a 15K would do 125).

applying math to our environment would mean up to 4800 100%random read IOPS on the DG, maybe reduced by overheads to 4000 IOPS

but "your" table would say max 2500 IOPS, however I can't understand why 100% read should be slower on VR5 than on VR1: it's striped accross all disks anyway

Maybe you can explain a bit ;-)
Marcus Longardt
Advisor

Re: safe IOPS

Thanks - but that isn't the point!!

I KNOW the max IOPS and I KNOW the IOPS which the system do! The vRAID, which kind of disc and how many of them.

What I don't know is how hp calculate the safIOPS thus the "amount of I/Os a diskgroup can do within a maximum of latency (15ms read/5ms write)" (see obove) is.

Does you know the forumla?

Regards
Marcus
S. Boetticher
Regular Advisor

Re: safe IOPS

I'd guess a mix from:
- physical info per spindle (accoring to RPM, disk-size platter and more you can calculate (or the manufacturer tells you) how many IOPS a HDD can do, then do the math
- real measurements: for instance http://h41112.www4.hp.com/promo/ess-trilogy-2/uk/en/pdf/HP%20StorageWorks%204400%20Enterprise%20Virtual%20Array%20(EVA4400)%20performance%20white%20paper.pdf


thus together would give the approximate "safe" numbers for good IOPS before latency goes above the limts...
Víctor Cespón
Honored Contributor

Re: safe IOPS

Marcus, I have the excel sheet with the formulas, but I can't share it here.

The amount of IOPS depends on the number of disks, type of disks, RAID type and % of writes. I can input your numbers and give you an estimate.

As you can see on the EVA4400 paper above, the latency grows slower with 100% reads.

The safeIOPS is a reference, a number that is very likely to give you low latencies. But I've seen many EVAperf captures and even if you are above that in some moments, the latencies keep being good.

S. Boetticher
Regular Advisor

Re: safe IOPS

@vcespon:

can I take your offer? please calculate a EVA4000 with 48x300GB 10K FC drives in 1 DG, 100% random read, only VRaid5

I know, that there is more than that, however if I know the max safeIOPS for my environment I can do some extrapolations, when we'll hit the perf-limits if we put even more stuff on it than now... ;-)
Víctor Cespón
Honored Contributor

Re: safe IOPS

Hi, that one is easy, 5,678 IOPS (118.3 per drive).

10% writes lower that to 4,353
20% to 3,525
40% to 2,548
...
100% writes to 1,374

As expected, 100% writes is a bad idea in a RAID 5, specially if using FATA disks. Unfortunately this is a situation often found in customers as they use FATA disks for backup and RAID5 to maximize space.