I ran some benchmarks comparing the Intel DC S3500 vs the Intel DC S3700 when being used as a SLOG/ZIL (ZFS Intent Log). Both SSDs are what Intel considers datacenter class and they are both very reasonably priced compared to some of the other enterprise class offerings.
Update: 2014-09-14 — I’ve done newer benchmarks on faster hardware here: https://b3n.org/ssd-zfs-zil-slog-benchmarks-intel-dc-s3700-intel-dc-s3500-seagate-600-pro-crucial-mx100-comparison/
Since flushing the cache to spindles is slow, ZFS uses the ZIL to safely commit random writes. The ZIL is never read from except in the case power is lost before “flushed” writes in memory have been written to the pool. So to make a decent SLOG/ZIL the SSD must be able to sustain a power loss in the middle of a write without losing or corrupting data. The ZIL translates random writes to sequential writes so it must be able to sustain fairly high throughput. I don’t think random write IOPS is as important but I’m sure it helps some. Generally a larger SSD is better because they tend to offer more throughput. I don’t have an unlimited budget so I got the 80GB S3500 and the 100GB S3700 but if you’re planning for some serious performance you may want to use a larger model, maybe around 200GB or even 400GB.
Specifications of SSDs Tested
Intel DC S3500 80GB
- Seq Read/Write: 340MBps/100MBs
- 70,000 / 7,000 4K Read/Write IOPS
- Endurance Rating: 45TB written
- Price: $113 at Amazon
Intel DC S3700 100GB
- Seq Read/Write: 500MBs/200MBs
- 75,000 / 19,000 4K Read/Write IOPS
- Endurance Rating: 1.83PB written (that is a lot of endurance).
- Price: $203 at Amazon
- Avoton C2750 (see build information here)
- 16GB ECC Memory
- VMware ESXi 5.5.0 build-1892794 (includes patch to fix NFS issues).
Virtual NAS Configuration
- FreeNAS 18.104.22.168 VM with 6GB memory and 2 cores. Using VMXNET3 network driver.
- RAID-Z is from VMDKs on 3×7200 Seagates.
- SLOG/ZIL device is a 16GB vmdk on the tested SSD.
- NFS dataset on pool is shared back to VMware. For more information on this concept see Napp-in-one.
- LZ4 Compression enabled on the pool.
- Encryption On
- Deduplication Off
- Sync=Standard (VMware requests a cache flush after every write so this is a very safe configuration).
Don’t try this at home: I should note that FreeNAS is not supported running as a VM guest, and as a general rule running ZFS off of VMDKs is discouraged. OmniOS would be better supported as a VM guest as long as the HBA is passed to the guest using VMDirectIO. The Avoton processor doesn’t support VT-d which is why I didn’t try to benchmark in that configuration.
Benchmarked Guest VM Configuration
- Benchmark vmdk is installed on the NFS datastore. A VM is installed on that vmdk running Ubuntu 14.04 LTS, 2 cores, 1GB memory. Para-virtual storage.
- OLTP tests run against MariaDB Server 5.5.37 (fork from MySQL).
I wiped out the zpool after every configuration change, and ran each test three times for each configuration and took the average (in almost all cases subsequent tests ran faster because the ZFS ARC was caching reads into memory so I was very careful to run the tests in the same order on each configuration. If I made a mistake I rebooted to clear the ARC). I am mostly concerned with testing random write performance so these benchmarks are more concerned with write IOPS than with throughput.
# sysbench –test=fileio –file-total-size=6G –file-test-mode=rndrw –max-time=300 run
# sysbench –test=fileio –file-total-size=6G –file-test-mode=rndwr –max-time=300 run
OLTP 2 threads:
# sysbench –test=oltp –oltp-table-size=1000000 –mysql-db=sbtest –mysql-user=root –mysql-password=test –num-threads=2 –max-time=60 run
OLTP 4 threads:
# sysbench –test=oltp –oltp-table-size=1000000 –mysql-db=sbtest –mysql-user=root –mysql-password=test –num-threads=4 –max-time=60 run
|none||OLTP 2 Threads||151|
|Intel DC 3500 80GB||OLTP 2 Threads||188|
|Intel DC 3700 100GB||OLTP 2 Threads||189|
|none||OLTP 4 Threads||207|
|Intel DC 3500 80GB||OLTP 4 Threads||271|
|Intel DC 3700 100GB||OLTP 4 Threads||266|
|Intel DC 3500 80GB||RNDRW||1120|
|Intel DC 3700 100GB||RNDRW||1166|
|Intel DC 3500 80GB||RNDWR||588|
|Intel DC 3700 100GB||RNDWR||569|
Surprisingly the Intel DC S3700 didn’t offer much of an advantage over the DC S3500. Real life workload results may vary but the Intel DC S3500 is probably the best performance per dollar for a SLOG device unless you’re concerned about write endurance in which case you’ll want to use the DC S3700.
There are a few SSDs with power loss protection that would also work. The Seagate 600 Pro SSD, and also for light workloads a consumer SSDs like the Crucial M500 and the Crucial MX100 would be decent candidates and still provide an advantage over running without a SLOG.
I ran a few tests comparing the VMXNET3 vs E1000 network adapter and there is a performance penalty for the E1000. This test was against the DC 3500.
|E1000g||OLTP 2 Threads||187|
|VMXNET3||OLTP 2 Threads||188|
|E1000g||OLTP 4 Threads||262|
|VMXNET3||OLTP 4 Threads||271|
I ran a few tests with Encryption on and off and found a small performance penalty for encryption. This test was against the DC S3700.
|Off||OLTP 2 Threads||195|
|On||OLTP 2 Threads||189|
|Off||OLTP 4 Threads||270|
|On||OLTP 4 Threads||266|