From Mark Klein of Orbit Software:

There are a couple of things to consider here when examining the trade-off between hardware and software compression. When software compression is on, less data is passed down the channel to the drive at the expense of CPU. When software compression is off and hardware compression is on, more data is passed down the channel, but nearly the same amount is finally put on tape.

Now, depending on your environment, if you're IO or CPU bound, the results of enabling hardware/software compression can significantly change performance. If you are neither IO nor

CPU bound, the effects of such a change between hardware and software compression should be minimal.

You effectively end up with the following matrix:

IO Bound CPU Bound Both IO and CPU Neither

(1) (2) (3) (4)

(1). Since you already have IO bandwidth problems, you want to minimize the amount of data sent to the backup device in this case. You would prefer software compression with no hardware compression on the target device. (A word of warning here: using high software compression with hardware compression enabled may actually cause the data to expand over what it would have been without any compression whatsoever. Do not do it!)

(2). Since software compression can require heavy doses of CPU depending on the method chosen, you would prefer to pump the data down to the device for compression. This keeps the CPU available for other tasks at the expense of greater IO.

(3). This one gets too tough to really discuss. There are a lot of trade-offs and switches with which some backup products can be tuned in such an environment. Each case tends to be different. Your software support people are best equipped to discuss this with you.

(4). In the case where you're neither CPU nor IO bound, the speed of the device will become the limiting factor and you will not see an appreciable difference between software or hardware compression.

[Editor’s note: The best thing to do is experiment with different combinations to find the right one for your environment. And then re-run your experiments every so often, but especially if you’ve made any significant hardware or software changes.]