sysbench slow performance of new 18 disk raid 1+0

We have recently purchased the following hardware.
HP DL380 8cores 32GB 2x72 10k sas drives
HP MSA70 array with 18 sas 146 10K dives

We upgrading our MYSQL database hardware.  I have the array configured  the array as a 18 disk raid 10 with 100 write cache on.

I am running the following sysbench test:

sysbench --test=fileio --init-rng=1 --file-total-size=8G --file-num=4 --max-time=120 --max-requests=10000000 --file-rw-ratio=1.5 --file-fsync-freq=on  --file-test-mode=rndwr --file-extra-flags=direct --num-threads=32 prepare

sysbench --test=fileio --init-rng=1 --file-total-size=8G --file-num=4 --max-time=120 --max-requests=10000000 --file-rw-ratio=1.5 --file-fsync-freq=on  --file-test-mode=rndwr --file-extra-flags=direct --num-threads=32 run

And getting the flowing results:
18disks raid10 Read 0b  Written 9.9037Gb  Total transferred 9.9037Gb (84.467Mb/sec)    5405.88 Requests/sec executed

I get 5400 random writes per second on the 18 disks raid 10.  Our dba is saying that he is getting 10K on a 4 disk raid10.  I believe  I am getting the correct performance out of this array. But I do not have a real base line.  I am looking for benchmarks of similar systems `16-20 disk raid 10 with the same sysbench command I ran above.

I am also looking for any configuration problems that would lead to 4disks running 2 times as fast as 18 disks.
Thank You

Who is Participating?
andyalderConnect With a Mentor Commented:
5400 IOPS sounds pretty good to me, 600 IOPS per disk so some is being met from cache - you'll get about 170 IOPS per disk at a push. Not sure how he's getting 10K IOPS on 4 disks, that's 5000 IOPS for a single disk which is clearly impossible so it has to be coming from RAM. It can't even be coming from controller battery backed RAM if you are using an 8GB file size since there aren't any PCI controllers with that much RAM on them.

Not sure how rndwr and file-rw-ratio act in combination, I guess it ignores file-rw-ratio.
martin_2110Author Commented:
Thanks andy. Do you know a good benchmark that will defeat caching? I would like a linux benchmark to give me the actual IOPS per drive?
martin_2110Author Commented:
Far as the file-rw-ratio it is ignored.
Improve Your Query Performance Tuning

In this FREE six-day email course, you'll learn from Janis Griffin, Database Performance Evangelist. She'll teach 12 steps that you can use to optimize your queries as much as possible and see measurable results in your work. Get started today!

Not really a Linux engineer but I presume you can do the same as I do using IOmeter for Windows. Partition the drive into a number of slices (say 20 of them) and then run 20 seperate instances of sysbench each using a file on its own partition. This defeats the cache fairly well because instead of an single 8GB file there are 20 * 8GB files spread across the disk. It also forces long seeks; your developer might be using big SATA disks in his PC and you can almost get 8GB on a single track on the outside of the platter. You'd obviously have to drop num-threads down to one per instance.

If he is using big SATAs then get him to run the tests for an hour or so, they go into write-verify mode when they get hot which really slows them down. Your little SAS disks are designed for 100% duty cycle so they don't slow down like that.
martin_2110Author Commented:
Thank you.
Thanks :)

I'd like to see the results of re-running the tests on the two arrays so will be grateful if you can post them sfter re-running. If nothing else it will help others because as you must have observed, there aren't many publically accessible benchmarks out there.
Question has a verified solution.

Are you are experiencing a similar issue? Get a personalized answer when you ask a related question.

Have a better answer? Share it in a comment.

All Courses

From novice to tech pro — start learning today.