Disk Enclosures
1748213 Members
2974 Online
108759 Solutions
New Discussion юеВ

Re: Does anyone know an estimated throughput to the Backplane from an Array (six 73gb 10k drives) using

 
SOLVED
Go to solution
Adam Hicks
Advisor

Does anyone know an estimated throughput to the Backplane from an Array (six 73gb 10k drives) using

Does anyone know an estimated throughput to the Backplane from an Array (six 73gb 10k drives) using dual channel, on an rp5470 w/two 875Mhz cpu's?
6 REPLIES 6
harry d brown jr
Honored Contributor

Re: Does anyone know an estimated throughput to the Backplane from an Array (six 73gb 10k drives) using

What kind of array? (SCSI or FIBRE)
What model of the server IO card?

live free or die
harry
Live Free or Die
Adam Hicks
Advisor

Re: Does anyone know an estimated throughput to the Backplane from an Array (six 73gb 10k drives) using

The Array is SCSI. It has an embedded I/O (a6696B) and also a Dual Port Ultra2 SCSI adapter (A5150A).
Tim D Fulford
Honored Contributor
Solution

Re: Does anyone know an estimated throughput to the Backplane from an Array (six 73gb 10k drives) using

The 10k drives (seagate ST373405LW, ST373307LC, ST373307FC) have similar seek times ~ about 8ms. The max transfer rate will vary 40 - 105 MB/s but this is not really that important as 8kB will take 0.2 - 0.07ms to transgfer

At 100% disk util using RANDOM access, you will get about 125 IO/s, each transfering 8kB, thus just under 6MB/s bandwidth

IF you are using a far higher block size, say 64kB, then this number will go up to about 40-43MB/s.

The above assumes you are JUST striping & there are no hotspot disks. If you are using RAID1+0 then halve the above! Also notice that the busses will probably not flood (unless you are using a SCSI less than Ultra160) and that you will probably not even push the rp5470's internals to any limit! the bottleneck is almost certainly the seek times of the disks.

HOWEVER....

at 100% disk util using SEQUENTIAL access (64kB) the figures get better, but are harder to calculate. The FC headders are 100Bytes per 2kB of payload I assume SCSI has similar overheads, so a 64kB block will require an extra 3kB. The track to trak seek is about 0.5 ms and the xfr time for 64kB + headder is about 1.5ms. so 2.0 ms total. This gives 3000 IO/s in total for all six disks or 187.5MB/s This is looking like a large fraction of 2x Ultra160, so you might not get it, but I'd imagine you should be able to get close to it say 160MB/s. At 160MB/s you will scan in all your disks in just under 8 minutes!

Unless you are doing massive sequential work the 187MB/s (160) figure is unreasonable high. It really depends on what you are using it for. I have an OLTP database & only get 10-15MB/s peak on 30 disks in RAID1+0 on a VA7400! (which is NOT unreasonable as the block size is so low!)

Regards

Tim
-
Eugeny Brychkov
Honored Contributor

Re: Does anyone know an estimated throughput to the Backplane from an Array (six 73gb 10k drives) using

Sorry for offtopic. Tim, did you have VA performance analysis done?
Eugeny
Tim D Fulford
Honored Contributor

Re: Does anyone know an estimated throughput to the Backplane from an Array (six 73gb 10k drives) using

Eugany

I'm not sure what you mean.

Do you mean have I ever had VA performance analysis on one of our VA's ... yes. If memory serves me well a chap called Duncan did it. It got a green light.


If you mean have I gone on a VA performance course. No, I've had no training for many a moon.

Regards

Tim
-
Tim D Fulford
Honored Contributor

Re: Does anyone know an estimated throughput to the Backplane from an Array (six 73gb 10k drives) using

Euganey..

I've just re-read my reply, I know know where you are comming from... My VA looks like it performs poorly...

The VA7400 deals with LOTS of small block IO. This is because the database is TRUELY OLTP. We do lots of 2.5kB IO's. The caching is 60% and I don't think we can improve it. Basically 100% write cache & 20% read caching (the read part SHOULD be going to the tables in the database memory, so if it is not, it is UNLIKELY to be in VA cache). The whole shooting match is very seek dominated. that said the caching helps enourmously. I'd expect that 30 disks in a JBOD would do 5.5 ms service time per mirrored pair, we get 2-4ms averaging 3ms. which is what you might expect for 60% cache. The throughput at 3ms service times is simple to calculate (& hence bandwidth), 3ms per LUN is 333 IO/s (1000/3), there are 13 LUNS RAID1+0 (we have 4 hot spare's) so 4330 IO/s for the VA, which is just over 10MB/s bandwidth.

I appriciate looking at the above number the VA looks poor, but compared to JBODs it is nearly twice as good because of the caching. The whole bottle neck is the back-end disks, NOT the FC, not the HBA not the controller. Admittidly If the block size was increased things would get better, but doing that we would only increase the ammount of un-used data read in (like having massive read-ahead which is unused). The way to improve things is to put more disk on it say 60 or 90, possibly more memory (VA7410 can go to 2GB). But at something like 10,000 ├в ┬м/GB it is expensive, you can buy alot of disk (performance) for 1GB memory.

Regards

-