Disk I/O performance issues...

[Richard Zeigler]

Greetings -

I have been battling what appears to be a disk I/O bottleneck. I was hoping that one of you smart people might be able to provide some insight into what appears to be an under-performance on DL320-G2 (3.0GHz) servers.

The following table represents some data that was collected using the PAT (Performance Assessment Tool).

Server Type Quick Sequential
99.21 DL-380-2400 15.3 MB/s 58.4 MB/s
97.21 DL-380-3000 19.6 MB/s 60.7 MB/s
99.23 DL-320-3000 9.4 MB/s 49.2 MB/s
97.22 DL-320-3000 9.5 MB/s 50.3 MB/s
97.23 DL-320-3000 10.7 MB/s 48.8 MB/s
99.22 DL-320-1000 5.7 MB/s 35.9 MB/s

The DL-320 servers are equipped with single ATA-100 drive (no RAID).
The DL-380 servers are equipped with dual SCSI Ultra 3 or 320 drives (RAID 1).

My questions are:

Should it be expected to have the DL-320 perform so much worse (30-50%) on disk performance?

Is this strictly attributable to the differences in IDE ATA-100 vs. SCSI SmartArray5i?

Doesn't ATA-100 imply 100 MB/s transfer rates?

Do these numbers look believable?

The drives/version are as follows:

DL-320: Integrated Ultra ATA-100 IDE RAID controller; LSI Logic Corp version 2.5.2003.613 (date 6/13/2003)

DL-380: SmartArray 5i; HP 5.62.0.32 (date 6/14/2004)

Thanks in advance for any assistance that will help my understanding or set the correct expectation.

Regards - Ziggy

P.S. This thread has been moved from Disk Array to ProLiant Servers (ML,DL,SL). - Hp Forum Moderator

[John Kufrovich]

Ziggy,

When looking at performance numbers you have to look at your intended use. There are two items of interest IOPS and MB/s. Way to many individuals concentrate on MB/s.

If your block size is small, you will have high IOPS but low MB/s. Databases are a good example of IOPS performance. SQL uses 8k blocks, Oracle uses either 8K or 16k.

If your block size is large, you will have low IOPS but High MB/s. Typically, backup apps, use 64k to 128k.

The industry will throw numbers at you. 30K IOPS per sec or 160MB/s. These are the max.
Take for instance 30K IOPS, usually this is very small data, 512B and it usually is just cache hits.
Like wise, 160MB/s is usually very large block transfers.

There are MANY other criteria that can impact performance numbers.

A good tool we use in our lab is IOMeter available free from IOMeter.org.

[Richard Zeigler]

Thank you John -

I can appreciate the IOPS vs MB/s issue from a quantitative standpoint. However, one of my bigger concerns is the relative difference in the measurements between the DL380 and DL320 servers. The 30-50% drop in MB/s caught me off guard (not the fact that it was such a small number).

All of these measurements were taken with the same OS image, no extraneous network activity, and the same applications running (none of which have any disk I/O).

Can you offer comment on the relative I/O behavior?

Thanks for considering it - Ziggy

[John Kufrovich]

I believe the 320-g2 are using SATA. Therefore they are capabable of 150MB/s.

The biggest item is
RPMs,
SATA are 7.2K rpm
SCSI are 10/15k

What are speeds of your SCSI drives?

jk

[Richard Zeigler]

Hi John -

The DL380 has 10K drives.

I was under the impression that the DL320 with the ATA-100 was limited to 100 MB/s. Is that incorrect?

Thanks - Zig

[John Kufrovich]

True,
There are several iterations of ATA
ATA 100 = 100MB/s
ATA 133 = 133MB/s
SATA 1.5Gb/s = 150MB/s

But realize that with ATA, the rpms could be either 5.4K or 7.2K.

[Vincent Fleming · Accepted solution]

I'm not surprised at your performance numbers at all - they're typical.

Here's why:

The RAID-1 is using a RAID controller. The controller will introduce latency and slow down transfer rates, particularly on writes.

With writes, both drives must be written to before the application (in this case PAT) is told that the I/O is complete. Depending on the implementation of the RAID, this can be as much as 2x the time it takes to write to one drive.

To get performance out of RAID, you need more drives to make up for the latency introduced by doing the RAID; this generally shows much better aggregate performance - not necessairly better single-threaded performance.

Cache can help, but not with sequential workloads - since you sustain the I/O for a long period, the cache fills, and the transfer rate slows down to what the underlying devices can handle.

If you want to see aggregate performance, you should use a benchmark that uses multiple threads, such as Iometer.

As for bus speeds, ATA vs SCSI, forget about bus speeds when using 1 or 2 drives - the bus speed means that the drive can send or receive BURSTS of up to that speed (in the SCSI it can burst up to 320 MB/s, for example), but the drive itself can only take the data off (or put it on) the platter at about 50-60MB/s at best.

So, you will NEVER sustain performance anywhere near the bus speed with only 1 drive on the bus. You would have to stripe the drives to get anywhere near the rated bus speeds.

Cheers,

Vince

[Richard Zeigler]

Thank you both for the information.

Vincent - I can definitely appreciate the slower writes on a RAID controller. However, the DL380 are the servers with the SCSI (10K RPM) RAID (faster), while the DL320 are the single drive IDE ATA/100 (7K RPM) (slower).

I will definitely follow the suggestion made by both of you and use the benchmarking tool.

Regards - Zig