What Sustained Disk Transfer Rates Do You Get?

Mr. Jackson asks: "What kind of disk transfer rates (MB/s) do people get in the real world when moving around large (100s MB) files? Either every machine in our building is mis-configured, or our notions about what we were getting are way off. I've tested half a dozen machines, mostly Win2k, some Linux, by just copying a large file and timing it with a watch. 8 MB/s seems to be about average for inter-disk copies. RAID 1 (stripped) got as high as 12 MB/s after fiddling with cache settings. RAID 5 was as low as 2 MB/s. We all thought the numbers should have been around 30 MB/s."

Comments

RAID intricacies

[photon317]

On RAID technologies, speaking in general terms assuming vendors do a good job of implementing it, here's a summary:

RAID 0: Pure striping, maximum performance, no redundancy. Cost is the same as concatenating disks to get the space you need.

RAID 1: Pure Mirroring, full redundancy - reads can be as fast as a stripe of the same width as the number of mirrors (2-way stripe, 2-way mirror, same read speed, etc) if they do round-robin reading. Writes happen in parallel, and can be slower unless you've got the headroom and the disk spindle is the only write bottleneck. Cost is double a simple concat or stripe.

RAID 2-4: Sometimes used for very special purposes, but generally ignored by all because one of the other raid levels does the same thing better. I've seen RAID-3 recently, there are occasionally valid uses for like 0.01% of people out there.

RAID 5: You get some data redundancy to survive a single disk failure, but you don't pay the double disk cost of full mirroring. It's an N+1 type of configuration. Speed is generally the slowest compared to everything else.

Now on top of those very basic things, there are other factors. Because RAID-5 is cheapest disk-wise, and (IMHO) because it has the highest number of the well-standardized RAID levels, RAID-5 is very popular. To make up for RAID-5's abysmal performance, people use hardware RAID-5 accelerators with cache and whatnot. The problem there is that the controller can add significant cost (in some cases enough to have paid for a full mirror in plainly controlled disks), and that the RAID controller itself can become a single point of failure.

At my office (where a lot of bad decisions get made every day and I have to eat it) they built a Veritas cluster of Sun machines around a SAN. The idea was that no node was a single point of failure because of clustering (with veritas allowing all nodes to reach the SAN storage). However, the SAN storage was a big fat RAID-5 array with redundant controllers/disks/yadda/yadda. Of course, as much as the vendor tries to bury it in the fine print, the RAID-5 hardware is a single point of failure. Sure enough, our very reputable vendor's "redundant" hardware raid-5 controller did fully fail one, knocking our data offline for hours.

For the same cost as the expensive raid-5 array and the disks in it, we could have bought two independant JBOD arrays (just a bunch of disks, no raid controller), placed them on the redundant SAN, with the redundant clustered machines doing software mirroring to the disks, and been truly free from single points of failure (assuming we do all the details right - that the mirrors are always across seperate arrays, and that the arrays are on seperate power, etc)..

I've spent a lot of time on these problems, and it is my strong belief that the optimal solution for almost all normal situations where you want high availability is to do software mirror/stripe (1+0). Be careful that there is a difference between 1+0 and 0+1 when the 0 part's stripe is more than two disks wide... Consider two JBOD arrays of 5x 36G disks each...

In 0+1, you first stripe each array into a 180G stripe, then mirror the two together. When your first disk fails, nothing so mcuh as hiccups. However, of your remaining 9 disks, if any of the 5 disks in the array opposite the one with the first disk fails, you will lose data. Thus there's a 5/9 chance that the second disk failure causes data loss.

In 1+0, you first mirror each disk from the first array with its partnet in the second array. You then take your 5 36G mirrors and stripe them together for your 180G. Again, first failure, no hiccups. If a second disk fails, in order to cause data loss it must be the partner of the first failed disk - any of the other disks can fail and you still lose nothing. So the chances of data loss on a second disk failure are now 1/9 instead of 5/9.