0

I have an OVZ node with 46 VPSes. The SSD drive had 400Mb/s but after filling the node with VPSes the SSD drive sometimes drops to 40Mb/s. How can I prevent this from happening? Is there a way to find and suspend users who have abused their IO?

user3180
  • 337
  • 1
  • 4
  • 13

2 Answers2

1

The SSD drive had 400Mb/s but after filling the node with VPSes the SSD drive sometimes drops to 40Mb/s.

Full SSD, running out of free sectors and triggering ome internal optimizations. Yes, happens. See, running ONE SSD simply may not be enough. Time to get some proper higher end stuff. Raid 10 of multiple SSD? Stuff like that.

Also, MB/S even for SSD is a bad metric. How many IOPS?

TomTom
  • 51,649
  • 7
  • 54
  • 136
  • Thanks, that's what I was thinking. Does putting them in software RAID1 usually solve this problem? Raid10 strains my budget. Would the 46 VPSes all be able to get to 400mb/s in a 2xSSD RAID1 setup. – user3180 Aug 25 '13 at 15:05
  • No idea - seriopusly, we live in different worlds, my servers run RAID over SAS with an enterprise grade RAID card. Double SSD wont likely give you anything unless the raid software distributes the writes, which is an implementation detail I am not sure it does. – TomTom Aug 25 '13 at 15:19
  • The thing is, I don't think my SSD is full. It's only using 258Gb out of 512Gb. Would reducing the disk space for each VPS help? What about RAID1 with hardware RAID? – user3180 Aug 25 '13 at 23:01
  • Dude, this is an IOPS problem, not a size problem. Arguing with "not full" is not going to help you. PRofessional admins (cough) of databases have many years bought a lot more size to get the needed IOPS budget. "Short Stroking" is standard in their world. Your 46 machines overload the SSD IO capabilities. Storage size is a totally different angle and not relevant. – TomTom Aug 26 '13 at 03:57
  • Ok, I got that now. So the only way to increase IO is to stick them into hardware RAID10? I'm confused as to what "running out of free sectors" means. – user3180 Aug 26 '13 at 07:43
  • Well, read about how SSD work. They do not overwrite - they allocate new sectors. Once they run out they mave have to reorganize. DUring this they may be slower. IMplementation detail - depends on the SSD - but yes, happens. And yes, you need more IO - get more hardware. As simple as it is. – TomTom Aug 26 '13 at 07:45
  • Would decreasing the number of users help? Or, is it possible that one abusive user use up all the IO – user3180 Aug 28 '13 at 15:27
  • Frankly speaking , the host should have an IO watch on that. At least I can see the IO usage of every file that represents a hard disc when I log into the server. And the scheudler should give you a fair amount still. This must be TERRRRRIBLE oversold. But yes, in general - moving users off is identical to increasing the available IO budget. – TomTom Aug 28 '13 at 15:30
  • Thanks my man. Whats a good io watch software? I also can't understand how to "short stroke". From some articles I read it seems like specific software is needed. How would I short stroke intel 520 cherry? – user3180 Aug 28 '13 at 21:39
  • You do not. As I wrote - short stroking is HARD DISC - and it simply means onlhy have a smalle rpartition. SSD do not work that way - no moving physical parts. And no, YOU do not use ANY software - the HOST has (the ony controlling the physical machine) and - I dont use Linux, I use only Hyper-V, so I wont recommend on specific software. – TomTom Aug 29 '13 at 05:27
-1

Consider running in a RAID 10 or RAID 0 setup to boost the speeds.

Do you have a control panel? That can typically help you figure out who an abuser is.

Jason
  • 3,931
  • 19
  • 66
  • 107