Understanding I/O performance

The disks are physically on the machine; they're in a RAID10, I believe they're 15K RPM SAS drives.

Buying a larger plan reduces contention (less nodes on a box), but it really comes down to how busy the box you're on is.

As with any VPS provider, you'll often see better results by scaling horizontally rather than vertically. For example, if you need better database performance with a VPS provider, you're better off building a cluster of smaller boxes rather than one big box.

It may also be worth noting that I believe that Linode - unlike some other providers - only puts people with the same plan on a single host. So it's not like if you get a larger 720 or 1440 that you might still get destroyed by a ton of 360s competing with you.

I/O contention can still be a problem, but you at least know that you're competing with fewer DomU's as you go up in plans, and that everyone you're competing with is using a similar plan to you.

I have no idea how much difference in practice this makes, but at least from my perspective I feel this approach is a more desirable arrangement than mixing and matching plans at various levels on the same host.

– David

What worries me is that the larger plans are used by people who couldn't get by on a 360, which means they might be doing lots of I/O or something. :P

@Guspaz:

I believe they're 15K RPM SAS drives.
No way they're 15K RPMs. They're 1 TB or larger, which means 7200 RPM at best.

@db3l:

It may also be worth noting that I believe that Linode - unlike some other providers - only puts people with the same plan on a single host.
Yes, that's correct.

@mnordhoff:

What worries me is that the larger plans are used by people who couldn't get by on a 360, which means they might be doing lots of I/O or something. :P
Absolutely. Although given that it doesn't take a large absolute number of DomU's - of any plan - to effectively saturate the I/O bandwidth, I'd probably still take the bet that the smaller total number of DomU's with the larger plans give better odds at having better I/O than not on average.

– David

Sure doesn't seem like 1TB or larger, with the amount of disk space we're allocated…

@mnordhoff:

No way they're 15K RPMs. They're 1 TB or larger, which means 7200 RPM at best.

They put what, 40x360 on a host? That's 640GB of total storage required, double that for redundancy, 1280.

15K RPM drives come at up to 600GB. I believe Linode uses four disks per machine in RAID10, giving you 2400GB of total storage, 1200GB usable.

So, there is absolutely no reason why they can't be 15K RPM.

Actually, you're both wrong. USB thumb drives don't spin.

@ljosa:

Does buying a larger plan result in better I/O performance?
We generally recommend it, for the reasons already pointed out. Although sometimes it can seem counterintuitive depending on circumstance, in our experience, it does work.

Yes, a bigger instance means less people on your host box, and less potential IO. But it's by no means a guarantee; bigger instances probably tend to use more IO, and since any one instance can cause serious IO contention, it's not really any guarantee that you'll be better off. You could go from a box with light 360 users to a box with heavy 720 users, and suddenly your IO performance is worse despite having half as many people on the box.

That's why scaling horizontally is safer; you're guaranteed to get more than you have now because you're adding new instances. It's a lot more work to get it right, and it would be wonderful if we could just expand our linodes to get more disk performance, but it doesn't work like that; until we move to SSDs where seek times don't matter and IO can be much better scheduled as a resource like CPU time can be, this is going to be the way it is.

Does anybody know if the drives are 15K or 7200 RPM?

Yes, someone knows if the drives are 15k or 7200 RPM (caker, at the very least).