The performance of a server system is affected by the speed of the backup device, the network speed, the amount of main memory, the disk speed, the CPU speed, and the number of CPUs. The factors affecting the performance of a client system are filesystem traversing, generation of data, data on multiple disks, CPU speed, and number of CPUs.
This chapter explains
guidelines for choosing a configuration
measuring server performance
measuring client performance
Several factors determine which NetWorker server configuration best suits your backup and recover needs. The configuration consists of the hardware and software, which includes tape drives, jukeboxes, client systems, and network connection.
This section provides a few simple rules that you can use to guide your choices, and focuses on backup, since backup requires far more server capacity than recover and occurs so much more often.
 | Note: Please keep in mind that these are guidelines, and actual performance may vary. |
The goal in selecting a configuration is to balance the different hardware and software limitations to achieve the overall data handling capabilities you require. Start by looking at the limits of the major NetWorker configuration components: tape drives, network connection, jukeboxes, clients, and the NetWorker server itself.
Tape drives have a fixed maximum data transfer rate that they can handle. Since NetWorker automatically spans multiple tapes, the total tape capacity is not as important as the data rate. Table 12-1 shows the data transfer rate for several tape drives.
Table 12-1. Tape Drives and Data Transfer Rates
Drive | Data Transfer Rate |
|---|---|
DAT | 200 KB/s |
250 KB/s | |
500 KB/s | |
1.25 MB/s |
Different clients can generate data at different rates and, even within a single client, different types of files can generate different data rates. For example, symbolic links require as much processing as large data files, but produce no data. Consequently, the data rate produced by a backup of a single client can vary quite a bit. The numbers listed below are considered average transfer rates for each client system. However, it is a good idea to run several clients simultaneously to help smooth out fluctuations in each client's data transfer rate. Table 12-2 shows the data transfer rates for several types of clients.
Table 12-2. Client Data Transfer Rates
Client: 1 Backup | Data Transfer Rate |
|---|---|
Silicon Graphics | 600 KB/s |
PC/DOS/SPX | 80 KB/s |
PC/DOS/TCP | 150 KB/s |
Sun™ SS2 | 200 KB/s |
IBM® RS/6000 | 300 KB/s |
Ethernet has an upper limit on bandwidth of about 1 MB per second, but in practice, most Ethernet networks are limited to about 800 KB per second between a set of clients and a single server. For higher performance, use FDDI. Table 12-3 shows the data transfer rates for several networks.
Table 12-3. Data Transfer Rates for Networks
Network | Data Transfer Rate |
|---|---|
Ethernet | 800 KB/s |
FDDI | 8 MB/s |
The server must be able to handle the load of network packets, data movement, and tape drives in order to achieve the rates listed above. Most of the work on the server side is in data movement, context switching, and interrupt handling. The performance of all of these functions improves as the number of CPUs increases.
Jukeboxes provide automatic loading and unloading of tapes or optical disks. This capability assists the administrator in two different ways. During nightly backups, NetWorker uses the jukebox to automatically switch to the next tape when a tape fills up. During recovers, NetWorker uses the jukebox to load all of the tapes needed for the recovery without operator intervention. Table 12-4 shows data transfer rates and capacities for several jukeboxes.
Table 12-4. Jukebox Data Transfer Rates and Capacities
Jukebox | Maximum Data Transfer Rate | Capacity | Average Data Transfer Rate |
|---|---|---|---|
480 KB/sec | 50 GB | 480 KB/sec | |
1.25 MB/sec | 70 GB | 1.25 MB/sec |
To determine the capacity requirements of a jukebox for a scheduled unattended backup, select the jukebox with a capacity large enough to handle the largest possible amount of backup data. For example, a full backup of 60 GB requires a DLT2700 jukebox or two EXB-10i autochangers.
To determine how much disk space you need for the online indexes for quick recovers, first do a rough calculation of the amount of data backed up in a single schedule period (for example, week, month, or quarter). Use the guidelines in Table 12-5 showing how much data is backed up with different levels of backup.
Table 12-5. Percent of Data Backed Up for Each Type of Backup
Level | Percent of Data Backed Up |
|---|---|
Full | 100% |
Level 1-9 | 25% |
Incremental | 10% |
For example, a monthly schedule that has one full backup on the first Sunday, a level 5 backup on other Sundays, and incrementals every other day looks like Table 12-6.
Table 12-6. Example of the Percent of Data Backed Up
Level | % of data backed up |
|---|---|
1 Full | 100% |
26 Incremental | 260% |
4 Level 5 | 100% |
Total | 460% |
This table illustrates that 460% of the total amount of data is backed up over the course of a month. For example, a total of 10 GB of client data backed up using this schedule would result in about 46 GB of data on tape per month.
Now assume that you have decided on a browse policy of two months for all the client systems and a retention policy of six months. These policies let your users quickly recover any file, and any version of a file that they had during the past two months. And with some effort you can recover for them files that they had any time during the past six months. So you need six months times 46 GB, which yields 276 GB of capacity.
In practice, you need a little extra jukebox capacity, since there are a small number of “unavailable” volumes as NetWorker must wait to recycle a tape until after all the save sets on that tape have expired.
Finally, remember to plan for growth in the number of your files. While sites differ in the rate at which their files are growing, a rule of thumb is that you should purchase a jukebox, or set of jukeboxes, with about 50% more capacity than your current requirement.
Site A has approximately 70 GB of data on two networks of 50 clients and wants to schedule full backups for all of their data in one night (12 hours). This equation calculates the required data transfer rate to achieve this goal.
70000 MB / 12 hours = 6700 MB / hour = 1660 KB / second |
To back up the data with a single NetWorker server, this configuration is suggested:
IRIX NetWorker with the TurboPak option
two DLT2700 jukeboxes containing DLT tape drives connected to the NetWorker server, with licenses for each
two Ethernet network interfaces
at least 1600 MB of free disk space on the server for the client index files
50 Client upgrade license
Site B has 50 GB of data on a single network with 80 clients; the administrators want to be able to schedule backups in a single night (eight hours). The full backups for the clients must be staggered due to the limit of 800 KB/sec data transfer rate per Ethernet network. Calculate the backup capacity required to complete the backups in one night:
800 KB / second * 8 hours = 23 GB / night |
By using three different backup schedules to stagger their full backups into three nights instead of one, the administrators reduce the load of the nightly backup data from 50 GB per night to about 20 GB per night.
Full: 50 GB / 3 = 16.7 GB Incr: (50 GB - 16.7 GB) * .1 = 3.3 GB Total: (16.7 GB + 3.3 GB) / 8 hours = 694 KB / second |
To back up the data with a single NetWorker server, this configuration is suggested:
IRIX NetWorker with the TurboPak option
one EXB-440 jukebox with two EXABYTE 8505 tape drives connected to the server, with license
one network interface
approximately 1 GB of free disk space on the server for client index files
50 client upgrade licenses
This section provides examples of how to measure the performance of a server.
Most tapes have a step function in their data rate. NetWorker uses 32 KB per record. To measure tape speed, follow these steps:
Create a large file (at least 20 MB) with non-zero data and list its size. For example:
# cat /unix /unix /unix /unix /unix /unix /unix /unix /unix /unix > big # ls -l big -rw-rw-r-- 1 root sys 20675420 Mar 5 16:11 big
Use the dd(1M) command to write the large file to tape four times and measure the time results:
# time dd if=big of=/dev/rmt/tps1d6nrnsv bs=32k conv=sync 95.2 real 13.0 user 11.9 sys # time dd if=big of=/dev/rmt/tps1d6nrnsv bs=32k conv=sync 78.2 real 12.9 user 12.7 sys # time dd if=big of=/dev/rmt/tps1d6nrnsv bs=32k conv=sync 78.0 real 12.8 user 12.5 sys # time dd if=big of=/dev/rmt/tps1d6nrnsv bs=32k conv=sync 76.8 real 13.0 user 12.4 sys
Divide the file's size by the average of the last three real times. For example:
Rate: 20675 KB / 77.66 seconds = 266 KB / second
This number gives you the rate of the tape speed.
NetWorker uses TCP and RPC/XDR as network communication protocols. To measure the network speed, follow these steps:
Create a large file (as in the tape speed measurement example) on a fast client.
# cat /unix /unix /unix /unix /unix /unix /unix /unix /unix /unix > big # ls -l big -rw-rw-r-- 1 root sys 20675420 Mar 5 16:11 big
Use the rcp(1C) command to copy the file from the client to the server and time the result:
# time rcp big server:/dev/null 38.2 real 0.2 user 30.7 sys
To find the network speed, divide the number of bytes in the file by the real time. For example:
Rate: 20675 KB / 38.2 seconds = 541 KB / second
The most important factor affecting network speed is network errors. To determine the input error rate, the output error rate, and the collision rate, use the netstat –i command. If the input or output error rate is above 0.5%, or the collision rate is above 5%, network errors are slowing down the network speed.
This section provides examples of how to measure the performance of a NetWorker client.
To measure the filesystem traversing speed, follow the steps below:
Time the uasm(1M) command with the –bi option. For example:
# time /usr/etc/uasm -bi /usr 33848 records 6961176 header bytes 644814472 data bytes real 51.86 user 7.73 sys 27.95 7.7u/27.9s (68% of 0:51) 0k+0k+0k 0pf+0sw 2993i+16o
Divide the number of records by real time for rate per file. For example:
33848 records / 51.86 seconds = 652.7 files / second
To measure the rate at which a client generates data for a backup, follow the steps below:
Time the uasm command with the –si option and redirect the output to /dev/null. For example:
# time /usr/etc/uasm -si /usr > /dev/null real 6:39.49 user 12.62 ys 1:27.19 12.6u/87.2s (24% of 6:39) 0k+0k+0k 0pf+0sw 45302i+27o
Divide the number of bytes obtained (filesystem traversing) with the uasm –bi command by the real time generated by the uasm –si command. For example:
629701 KB / 399.49 = 1576 KB / second
NetWorker automatically backs up multiple disks in parallel.
To measure parallel disk speeds, follow these steps:
Use the df(1) or du(1M) command to find two directories of approximately the same size on different disks.
Run the same uasm speed tests for filesystem traversing and data generation rate as for one disk, but run the tests simultaneously on the two directories.
Add the data from each test (files/sec and KB/sec) to obtain a combined rate.
This rate reflects the performance of NetWorker backing up data on multiple disks.