We ran unixbench 5.1.2 on one of our servers. We wanted to see how it compared to some of the competition. The Linode and EC2 benchmarks were done back in December 2009, on the latest hardware that was provided. Our OpenSolaris benchmarks were done on a system we bought back in March 2009.
We'll start off with Linode 360 plan (4 CPU / 4 parallel):
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 29911700.2 2563.1
Double-Precision Whetstone 55.0 7852.9 1427.8
Execl Throughput 43.0 5470.0 1272.1
File Copy 1024 bufsize 2000 maxblocks 3960.0 315110.5 795.7
File Copy 256 bufsize 500 maxblocks 1655.0 82099.9 496.1
File Copy 4096 bufsize 8000 maxblocks 5800.0 866155.2 1493.4
Pipe Throughput 12440.0 2053207.3 1650.5
Pipe-based Context Switching 4000.0 237263.9 593.2
Process Creation 126.0 10784.4 855.9
Shell Scripts (1 concurrent) 42.4 9259.1 2183.7
Shell Scripts (8 concurrent) 6.0 1539.9 2566.5
System Call Overhead 15000.0 1768915.5 1179.3
========
System Benchmarks Index Score 1254.5
Now, Amazon's EC2 medium (2 CPU / 4 parallel):
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 24194215.0 2073.2
Double-Precision Whetstone 55.0 8422.0 1531.3
Execl Throughput 43.0 2379.7 553.4
File Copy 1024 bufsize 2000 maxblocks 3960.0 142163.5 359.0
File Copy 256 bufsize 500 maxblocks 1655.0 36551.4 220.9
File Copy 4096 bufsize 8000 maxblocks 5800.0 421398.5 726.5
Pipe Throughput 12440.0 239183.7 192.3
Pipe-based Context Switching 4000.0 82291.4 205.7
Process Creation 126.0 2974.6 236.1
Shell Scripts (1 concurrent) 42.4 5357.8 1263.6
Shell Scripts (8 concurrent) 6.0 737.7 1229.5
System Call Overhead 15000.0 1467331.5 978.2
========
System Benchmarks Index Score 579.6
Finally, our OpenSolaris VPS servers with about 20-30% utilization. The first one is with 2 CPU / 2 parallel:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 54106806.1 4636.4
Double-Precision Whetstone 55.0 7594.6 1380.8
Execl Throughput 43.0 1849.9 430.2
File Copy 1024 bufsize 2000 maxblocks 3960.0 126718.2 320.0
File Copy 256 bufsize 500 maxblocks 1655.0 32468.1 196.2
File Copy 4096 bufsize 8000 maxblocks 5800.0 378133.6 652.0
Pipe Throughput 12440.0 2062538.7 1658.0
Pipe-based Context Switching 4000.0 262259.5 655.6
Process Creation 126.0 2888.7 229.3
Shell Scripts (1 concurrent) 42.4 4155.4 980.0
Shell Scripts (8 concurrent) 6.0 728.6 1214.4
System Call Overhead 15000.0 930953.7 620.6
========
System Benchmarks Index Score 724.0
We close out with 4 CPU / 4 parallel test on the same server:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 101277172.7 8678.4
Double-Precision Whetstone 55.0 15197.7 2763.2
Execl Throughput 43.0 2764.1 642.8
File Copy 1024 bufsize 2000 maxblocks 3960.0 152318.4 384.6
File Copy 256 bufsize 500 maxblocks 1655.0 39096.9 236.2
File Copy 4096 bufsize 8000 maxblocks 5800.0 526719.9 908.1
Pipe Throughput 12440.0 4177112.1 3357.8
Pipe-based Context Switching 4000.0 650505.1 1626.3
Process Creation 126.0 4924.3 390.8
Shell Scripts (1 concurrent) 42.4 5666.8 1336.5
Shell Scripts (8 concurrent) 6.0 806.4 1344.0
System Call Overhead 15000.0 728384.9 485.6
========
System Benchmarks Index Score 1074.5
Clearly, we were not leaders in the benchmarks. But there is something interesting to be said about these results. Our OpenSolaris VPS servers make use of ZFS. ZFS does some interesting things. The File Copy tests lag behind greatly for one big reason. ZFS on our OpenSolaris VPS ensures reliability over performance. Let us explain.
Disks have cache on-board to help with speed. Most other operating systems and file systems will "tell" the application that a write is complete soon after the data is written to the disk cache (a temporary location, before it is written to the slower disk spindles). If the power were to fail at this point, the data is lost. Your business critical database thinks it wrote data to disk but in reality, it did not. It wrote to the cache. This leads to corruption and bad data.
ZFS doesn't do this. ZFS ensures that all data is properly flushed to disk on each commit. ZFS takes the safer and reliable approach to disk management. You can expect your database to be in good order if the system were to crash or if we were to ever experience a catastrophic power outage.
If we were to put aside the File Copy tests, we think we would have a pretty good chance of beating out the competition! In the future we'll run these same tests on a ZFS system with flush on commit turned off. But don't count on that being disabled on a production system. We prefer reliability over performance. Adding SSD to the mix should also improve things greatly, but you'll have to wait a little longer to experience OpenSolaris VPS with SSD flash disks.
