If you have more than one state-of-the-art tape drive behind your LAN-based backup server, chances are you have too many. That's right. If you've got a LAN-based backup server, and behind it is more than one tape drive capable of 45 MB to 50 MB, then you should probably rethink your design. Let me explain.
It has to do with the nature of streaming tape drives. To fully realize their advertised throughput, you must supply these tape drives with a constant stream of data equivalent to their native throughput times, no matter what compression ratio you're getting. Sending data at a slower rate will actually cause it to write more slowly than the speed of the stream of data you're sending. The result in many backup systems is an array of tape drives that are writing at a fraction of their advertised speed. This has caused many customers to question the validity of the drives' advertised rates, when they should be looking at their design.
How to figure it out
When addressing this issue, there are three key numbers to consider: the drive's advertised native throughput, its capacity and your data's compression ratio. The Web site http://www.storagemountain.com/hardware-drives.html lists all popular drive types and their advertised native throughput rates. To determine your compression ratio, use your backup software product to determine how much data is being stored on tapes that it has marked "full." Divide the average size of your full tapes by the native capacity of the drive (also available at www.storagemountain.com, and you've got your compression ratio.
For example, if you've got a 9940B tape drive, its native throughput is 30 MB. If you're storing an average of 300 GB on its 200 GB tapes, then your compression ratio is 1.5:1. If you multiply 30 MB by a compression ratio of 1.5, you get a target throughput rate of 45 MB. If designed properly, your backup system should be capable of reaching this throughput rate on that tape drive in your environment. Most streaming tape drives have two speeds: 0 MB (stop) and the target throughput rate. Consider the 9940B above. It can only write at 45 MB in an environment with a 1.5:1 compression ratio. As soon as you drop below 45 MB, the buffer that is being used to hold data before it goes to tape will be emptied before it is filled up again.
When that happens, the tape drive does not slow down. It keeps writing at 45 MB until the buffer is empty. Then it stops, rewinds and waits for the buffer to be full again before repositioning and starting up another 45 MB stream. If you're supplying less than 45 MB, it will again empty the buffer, stop, rewind and reposition.
Shoe-shining: A waste of time
The lower the throughput you're supplying, the more time the drive spends stopping, rewinding and repositioning - referred to as shoe-shining. If you're sending it something like 10 MB, it's spending the majority of its time shoe-shining and very little of its time actually writing. The result is that it doesn't write 10 MB; it writes 5 MB. Even if you're giving it 20 MB, it might only be able to write at 15 MB. You get the picture. Therefore, it's important to send a stream of data equivalent to the target throughput of the tape drive, which is defined as its native throughput multiplied by your average compression ratio.
This can be accomplished in many ways, the most common of which is multiplexing. With multiplexing, you back up several file systems from several clients simultaneously and send them all to the same tape drive. The backup software interleaves, or multiplexes, all of these backups to the tape drive at the same time, resulting in a data stream of sufficient speed to stream the tape drive. While this does increase the speed of the backup, it also negatively affects the speed of an individual restore from that backup, as the backup software must read the entire backup and throw out what it doesn't need for that restore. Please note, however, that I'm talking about a traditional LAN-based architecture where dozens or hundreds of clients are backed up across the LAN to a Gigabit Ethernet-connected backup server with tape drives behind it. Unless you're using a TCP offload engine (TOE) card, you are probably only going to get about 50 MB of throughput into that backup server and out to tape.
A few years ago, when Gigabit Ethernet first came out and tape drives were 5 MB, it made sense to put five to 10 tape drives behind each server. Now Gigabit Ethernet isn't any faster, but the tape drives are six times faster. The result is that it doesn't make sense to put two tape drives behind a server that can only stream one. Thus, my new favorite design: one server to one tape drive.
Variables to consider
Yes, there are some other variables that change these numbers, such as disk staging (where backups are staged to disk first and then copied to tape) and tape twinning (where one backup is sent simultaneously to two drives). TSM has been doing disk staging for years, and NetBackup and NetWorker now support it as well. The theory is that copying locally from disk to tape is faster than backing up to tape across the network. Tape twinning allows you to create an original and a copy at the same time, at the same speed as making one copy, thus making use of two tape drives behind a single 50 MB connection.
My point is simply that you should consider how many tape drives you're putting behind your backup servers and determine if you've bought (or are buying) too many for your server to handle. Believe it or not, you can actually do more with less.
W. Curtis Preston is president and CEO of The Storage Group, which provides consulting services to end users and analyst services to vendors.