SCSI Hardware RAID Performance.

[Shadow07]

I have been searching the net for any and all documents on SCSI Hardware RAID performance. There is a debate going on between me and a few techs on which RAID provides better performance when dealing with a multi-user system (i.e. Terminal Server). The two RAID technologies we are discussing are RAID1 and RAID5. Since the servers that we recommend, build out, and configure are 1U servers (can only support at most 3 drives), RAID1 and RAID5 are the only options.

If anyone has any links, documents or $0.02's, please reply.

Thanks for your time,

 

[alrox]

This totally depends on the controller for the most part. I assume you'll have ultra160 1 channel controllers with X amount of cache on them. RAID 1 shouldn't see any slowdown with hardware raid, and there's no point to RAID 5 with a max of 3 disks, you'd end up with the same amount of disk space and have the same level of redundancy, you'd need 2 drives to fail in each instance for the raid device to fail completely. Either setup will be able to read/write > 100megabits/second which is most likely the network you'll be putting them on.

 

[Shadow07]

Thanks for your reply. This is already been stated in our conversations. I wanted to know if anyone has seen any statistical data or reports between the two RAID array types.

I agree it all depends on the controller, but we are not discussing that. Obviously, you will have better performance when using Ultra160 over Ultra2. I wanted to know what RAID type provides better performance, period.

Thanks,

 

[HypNoTic]

The biggest advantage of RAID is parallelism. Using 3 drives in RAID 5 will give about the same performance as 3 drives in RAID 1 but a add few drives to your RAID 5 and see the difference. Using 10 drives in native hardware RAID 5 will give you about twice the performance of 10 drives in RAID 1. Also, keep in mind that 10 drives in RAID 5 require more "hardware" ressources than RAID 1 that will simply send the data to 2 drives instead of computing the parity of the array but will give you much more disk space. That's why a good hardware HBA is required when you need performance. Forget the software emulation, it's good for home but it's a non-sense for enterprise-wise administrator.

Thanks Google, here's a few links :
This is what you were looking for (PDF)
RAID Performance Issues

-HypNoTic

 

[Shadow07]

This would make sense, but in a multi-user environment (i.e. a Terminal Server running Terminal Services/MetaFrame) we use 1U servers that have at most 3 drive bays. Now, I don't know of any hardware RAID HBA that can configure a RAID1 array with 3 drives (using all three drives for data). With drive read activity consuming about 80% of the normal operation, it would make sense that a RAID1 array would provide better performance. Keep in mind that this is not a file server, nor a database server. If I were, then that 80% drops to about 50-55% of the normal operation.

Also, you need to keep in mind that if a drive were to fail, the performance of a RAID5 array drops about 20%, due to the RAID card computing the parity information to generate the missing data that resides on the failed drive. With RAID1, you won't experience a degridation in performance when you have a failed drive. You will experience a degrade in performance when the array is re-building, and that applies to either a RAID5 or RAID1 array.

Thanks, and keep them coming.

 

[SCSIRAID]

IBM's ServeRAID controllers support a mode called RAID 1E which is a striped mirror. This allows mirroring operation with more than two drives including odd numbers.

As to the relative performance between a 2 drive RAID 1 and a 3 drive RAID 5... In a small block random workload with mostly reads (usually 70%)... the RAID 5 will perform better simply because there are more drives to distribute the IO's to. The RAID 5 write penalty is hands down worse than RAID 1 but writes are only 30% of the workload. If you up the blocksize and go more sequential then the parallelism of RAID 5's striping again give it the edge on a 2 drive RAID 1 since the drives are going in parallel. The RAID 1 algorithm could exploit parallelism and play stripe games to improve sequential data rate. Even if it did, in this case it would be 2 in parallel vs 3 so advantage RAID 5.

If you go to 3 drive RAID 1E vs 3 drive RAID 5 then things flip. The 1E write penalty is less than RAID 5. Since the mirror is striped then Sequential Data Rate capability is the same and the distributed IO capability is the same (since both have the same number of drives). Equal read performance plus superior write performance means advantage RAID 1E.

As with all synthetic workloads.... they dont necessarily represent the real world.... so... YMMV.