Disk-to-Disk backup slower then disk-to-LTO

[Mark Janvier]

Hello all,

Recently I have been implementing disk-to-disk backup solutions in a SAN environment with several customers.
They have an EVA3000 as primary storage, a MSL6030 library with LTO-2 tape drives and a MSA1500Cs with SATA drives (MSA-20 shelves)as backup devices, all SAN attached.
I have noticed that writing a backup from the same server of the same data takes longer to write to the MSA then to the LTO drive. We Use Data Protector as backup software.

Can anybody give me an explanation why this is? Disk-to-disk backup should be faster, no?

[Duncan Edmonstone]

Why should disk to disk backup be faster?

The difference between disk and tape IO performance can be characterised as follows:

Disk - excellent random read & write performance, average sequential read & write performance.

Tape - excellent sequential read & write performance, poor random read & write performance.

So for a backup (which is effectively a sequential write) it's quite possible for a tape (particularly a LTO2) to be faster than disk.

The difference is when you come to restore a file from your backup - now seek times become important and as disk is great at random IO, it has low seek times and can restore the file much quicker than the tape (mainly because it can find the file much quicker than the tape)

HTH

Duncan

[Mark Janvier]

Disks should be faster because SATA disks should have a troughput of +/-180 MB/s and LTO-2 +/-30 MB/s.

[Duncan Edmonstone]

First off where are you getting these figures for SATA speeds? Most SATA drives have an average sustained transfer rate of < 58MB/s AND thats a hardware speed rather than a real practical speed. Also the LTO2 speed you are quoting is native only (not taking into account any compression).

Second are you actually writing to a NTFS filesystem on the ATA disks? Again filesystems are not especially designed for sequential IO performance and can slow things down.

HTH

Duncan

[Mark Janvier]

I got the spec from the SATA drive spec for proliant SATA drives. It states 1,5Gb/s if my math is correct that is +/- 180 MB/s. For desktop SATA drives it states 150 MB/s. The backup is not to 1 drive but an array of SATA drives behind an MSA-1500 controller. So an array of SATA drives should still(theoretically) outperform a single LTO-2 drive.
On the other hand, I completely agree with you that the difference between a tape backup and a disk backup is the file-system (or the lack of it) on the target. Do you think this could influence performance? Now we use NTFS. Could FAT32 speed up things?

[Duncan Edmonstone]

As I understand it the 1.5Gb/s is baiscally the speed from SATA controller to drive cache (standard marketing BS lists the specs in the way that makes them look faster than they really are - all the drive manufacturers are guilty of this) - thats patently *not* the speed you will get from a drive - you are right in that if enough drives are used in a raid set you should achieve better performance eventually than to tape, but then how many tapes were you writing to? Also what concurrency settings are you using in DP, and how many streams of data are you writing to the device - are you using the new DP5.5 file depot device type or just the old file type...

HTH

Duncan

[Mark Janvier]

We were writing to 1 LTO-2 device with standard concurrency settings (min 1 max 5). We were writing 1 stream. We use DP 5.1 old file type. In the mean time we have done some extra testing and what we see is that performance between a disk to disk copy or a disk to disk backup is almost the same and that the disk to tape backup is slightly faster (but in the same order). So the loss in performance we have in disk-to-disk compared to disk-to-tape almost maches the file system overhead (certainly when we take into account the overall CPU load on the server). We see the difference in performance grow if the data set used contains larger files. Which reflects perfectly your remark about the NTFS file system not performing good at sequential reads/writes. For me it now has been proven that the difference in performance we were experiencing was the fact that disk-to-disk backup uses 2 file system accesses and disk-to-tape only 1. Thanks a lot for the input.

[Mark Janvier]

As mentioned, the difference is the second file system access.