Been too long in coming, more MySQL SSD benchmarks!!!

I presented these at an internal MySQL professional services meeting about a month ago… its mostly a hodge podge of various benchmarks… but enjoy! The big difference in these benchmarks vs the other benchmarks is I am testing on the memoright GT drive, which is supposed to be one of the fastest SLC drives out their currently. Lets get right too it:

Looking at sysbench Random read/write iops:

R/W 1 Raptor 1 Mtron 1 Memoright
5000/5000 172 200 284
6670/3330 164 282 412
7500/2500 159 388 512
8000/2000 165 516 607
8333/1777 176 518 741
10000/0 161 5263 3844
0/10000 200 100 160

The Memoright drive is noticeably faster then the Mtron drive in the synthetic random read/write tests. But the 100% write tests are substantially lower then the mtron drive which is just weird.

In the sysbench OLTP test, the memoright was slower then the mtron I tested earlier:

OLTP TRANS PS:

  Raptor Mtron Memoright
trans ps 25 132 88

The generic sysbench #’s would make me think that the memoright should perform better, but its not.

DBT2 tests are a bit more realistic (sysbench’s oltp uses a single table), here are the numbers for DBT2:

DBT2 NEW ORDER TPM:

  Raptor MTRON Memoright
DBT2 Score 625 4900 4407

Memoright is still a tad slower… which is puzzling…

So Lets look at the Orion Numbers and see if anything shakes out:

ORION IOPS:

Threads 1 2 3 4 5
Raptor ( 16K 100% rd) 140 139 138 139 143
Mtron (16K 100% rd) 4555 4883 4886 4886 4881
Memoright 3449 3641 3642 3640 3639
           
Threads 1 2 3 4 5
           
Mtron (16K 50/50 rd/wr) 214 207 206 206 205
Raptor (16K 50/50 rd/wr) 173 171 174 174 178
Memoright 372 366 366 367 365
           
Threads 1 2 3 4 5
Raptor (16K 80/20 rd/wr) 153 149 152 155 155
Mtron (16K 80/20 rd/wr) 546 524 518 516 515
Memoright 823 828 824 828 828
           
Threads 1 2 3 4 5
Raptor (16K 90/10 rd/wr) 148 147 149 152 152
Mtron (16K 90/10 rd/wr) 1028 989 987 980 976
Memoright 1489 1475 1473 1484 1482

Here the memoright once again faster… so far the generic tests show the memoright clearly with the performance lead, but the “real world” type tests say something differently. This may have to do with the workload, or maybe the concurrency of the generic tests -vs- the database tests. Or maybe it has more to do with the mtron being faster on reads:

ORION Disk latency (MS) :

# threads 1 2 3 4 5 6 7 8 9 10
Mtron (16K 100% rd) 0.22 0.41 0.61 0.82 1.02 1.23 1.43 1.64 1.84 2.05
Memoright 0.29 0.55 0.82 1.1 1.37 1.65 1.92 2.2 2.48 2.75
Raptor ( 16K 100% rd) 7.11 14.34 21.62 28.71 34.96 40.96 46.04 51.21 57.12 62.75

Not really sure… one theory is this could have to do with skewed r/w #’s in the database tests. The generic tests are pure read/write from disk. An update has to perform multiple read/write requests. Its possible IO stat would should this. I will try and look at the read/write % in iostat next time I run benchmarks instead of relying on insert/update or transaction #’s… remember just a theory. It could also be some other oddity in my tests I did not account for… it was 6 months between the mtron tests and the memoright tests and I did not have an mtron to rerun the tests with… but I am getting off topic…

More Orion #’s take a look at the latency on the drives:

ORION Disk latency (MS) :

# threads 1 2 3 4 5
Mtron (16K 50/50 rd/wr) 4.65 9.62 14.53 19.4 24.34
Memoright 2.68 5.46 8.19 10.87 13.66
Raptor (16K 50/50 rd/wr) 5.78 11.64 17.21 22.98 28
           
# threads 1 2 3 4 5
Mtron (16K 90/10 rd/wr) 0.97 2.02 3.04 4.08 5.12
Memoright 0.67 1.35 2.03 2.69 3.37
Raptor (16K 90/10 rd/wr) 6.72 13.55 20.07 26.24 32.72

So what does this all mean? Well both the Memoright and Mtron drives are fast! Depending on the workload one drive maybe faster then another. None of my tests were particularly write Intesive, which is where the Memoright’s advantage lies. These drives will not replace memory anytime soon.

A really nice way to sum this up take a look here, just for grins I had another consultant ( Yoshinori Thanks! ) run the same DB2 tests on a server he was testing on. Their was no comparison in hardware, my little crappy 4 core ancient intel server with a single SSD, versus a nice 8 core machine with 8 disks… mysql versions, configurations, memory allocations, etc were all the same. Its still not scientific, so don’t flame me with I should have tested with the same hardware… I know! Basically we took a server you would buy today, and put it against a server that is unfit for anyone but me:) We were trying to make a mostly “IO” bound workload, but still you would think a brand new shiney server with good disk, should kill my server. Here are the dbt2 results:

DBT2 New Order TPM:

Raid 5 Raid 10 Matt’s Mtron Matt’s Memoright
4579 6139 4900 4156
8-disks 8-disks 1 disk 1 disk

WOW! Depending on the drive the SSD actually out performed 8 (10K SAS) drives in RAID 5. This is one test mind you… I am not saying that SSD can beat 8 disks in a RAID 5 all the time, just in our one set of tests. In fact if you start allocating more memory to the database, our shiney new box kicks the crap out of my old box. I wonder how the shiney new box would perfom if I put the SSD drive in it? Thats for another time. Also to be fair I did compare RAID 10 numbers from the shiney new box used above, to another vendors shiney new box… the one above was slightly faster then the other vendors shiney new box… so this is not the case of getting a lemon either.

Just wish I had more time to play! This stuff is sooooo much fun:)

This entry was posted in hardware, mysql, performance. Bookmark the permalink.

Comments are closed.