4 ms·
Here is my $5 digitalocean slice (created a few months ago): # dd bs=1M count=1024 if=/dev/zero of=test conv=fdatasync 1024+0 records in 1024+0 records o
by growt 13y ago
Here is my $5 digitalocean slice (created a few months ago):
# dd bs=1M count=1024 if=/dev/zero of=test conv=fdatasync
1024+0 records in
1024+0 records out
1073741824 bytes (1.1 GB) copied, 3.7375 s, 287 MB/s
# hdparm -tT /dev/disk/by-label/DOROOT
/dev/disk/by-label/DOROOT:
Timing cached reads: 16394 MB in 2.00 seconds = 8205.62 MB/sec
Timing buffered disk reads: 868 MB in 3.00 seconds = 289.17 MB/sec
- saltcod 13y agoJust ran this on my 1GB droplet at Digital Ocean: $ dd bs=1M count=1024 if=/dev/zero of=test conv=fdatasync 1024+0 records in 1024+0 records out 1073741824 bytes (1.1 GB) copied, 5.83509 s, 184 MB/s
- drzaiusapelord 13y ago287 MB/s vs 816 MB/s is huge. Linode is really doing a good job here. I have 4 VMs with them and am very happy. Can't wait to make this move.
- chacham15 13y agoParallelizability isnt always possible, but you should normalize the speed you get per dollar. 287 MB/sec / $5 ~ 57 MB/sec/dollar 816 MB/sec / $20 ~ 40 MB/sec/dollar
- jgillich 13y ago287 MB/sec / $20 ~ 14 MB/sec/dollar I don't think MB/sec/dollar makes sense simply because disk speed does not scale. If you need a (very) fast disk, Linode is obviously the better choice.
- TylerE 13y agoIf the difference matters, a $5 DO node likely isn't cutting it.
- chacham15 13y agoThats not the issue. The issue is if (and only if) you have a perfectly parallelizable task, you would be better off getting 4 x $5 DO servers with a rate of (4 x 287) 1148 MB/sec which is much better than 816 MB/sec for the same price. Or you could even get 3 $5 servers and have a rate of (3 x 287) 861 MB/sec which is still better and also cheaper.
- lunixbochs 13y agoThat's only when comparing $5 to $20. Does the $20 DO model also have 287MB/s?
- weixiyen 13y agoIt's not like the $20 DO instance is going to be faster
- msumpter 13y agoI think these new nodes are still pretty lightly used, give it a few months and then do a comparison. Benchmarking VPS/cloud instances can be tricky since you might be on a quite node or a heavily utilized node. You really need to provision many machines and run the checks at random times over a period and then average the scores out. I upgraded some of my VMs and they all failed to boot on their own and I had to manually boot each one after the migration.
- diorray 13y agoHere's my DO results (AMS1). This server was busy and we're actively using it for 4+ months [root@server ~]# dd bs=1M count=1024 if=/dev/zero of=test conv=fdatasync 1024+0 records in 1024+0 records out 1073741824 bytes (1.1 GB) copied, 6.29244 s, 171 MB/s
- saltcod 13y agoAnd here's a run I did on a shared host at A Small Orange for fun: 1024+0 records in 1024+0 records out 1073741824 bytes (1.1 GB) copied, 62.9023 s, 17.1 MB/s