I have a bit of discrepancy in FC read performance when compared to write performance. I setup a single FC target using a QLA2462. The backend storage is a 3ware 9650SE card with 12 SATA 750GB drives in a 12 drive RAID 5 with 256k Chunk. The initiator (Cent OS 5.2) saw the target as a disk without issue.
I began doing some performance testing using dd;
time dd if=/dev/zero of=/dev/sdb bs=1M count=16384 for writes
time dd if=/dev/sdb of=/dev/null bs=1M count=16384 for reads
basically reading and writing a 16GB file
I ran each 10 times my average read performance was 142.2 MB/s however write performance was 394 MB/s. Note this is raw device performance, no filesystem.
After noting this i created a file system (ext3), mounted the drive, and re-ran the tests
time dd if=/dev/zero of=/mnt/opene/testfile bs=1M count=16384 for writes
time dd if=/mnt/opene/testfile of=/dev/null bs=1M count=16384 for reads
This time my average read was 166 MB/s but my write performance dropped to 222.7 MB/s.
Anyone else see this kind of weird performance using FC target?
Well, I'm not 100% sure if this is why you are having the result, but if you are reading from a drive, you actually have to find the data written to it (unless your cache is warm, but it usually isn't). If you write, the disk or the raid card or the open-e target system itself may buffer your write. With a 16GB test file, you might have half of that cached by the open-e system itself, and parts of the rest of it cached by the raid card and the disks. Also, by being able to cache more of it, the system might be able to more efficiently stream data to the disks.
We have deployed systems with 16GB or over of ram that can be used for caching, so it may be cached. I don't know how much ram you have.
Try the same thing with a larger test file. Also, try using different blocksizes in dd.
I figured caching would affect the system which is why i used 16GB files for this test. The server only has 2GB of system memory and the 3ware card has 256MB or 512MB (not sure at this moment 9650SE).
I'm just trying to get some baseline performance so I know what to expect if we ever chose to actually deploy and Open-E system using fibre channel targets. Note his was done with version 5.X, has any additional work been done that may improve performance in 6.X?
FC Target Speed
Reply With Quote
