Transaction throughput

Generally, the speed of a database system is measured by the transaction throughput, expressed as a number of transactions per second. The two gating factors for Berkeley DB performance in a transactional system are usually the underlying database files and the log file. Both are factors because they require disk I/O, which is slow relative to other system resources such as CPU.

In the worst-case scenario:

This means that for each transaction, Berkeley DB is potentially performing several filesystem operations:

There are a number of ways to increase transactional throughput, all of which attempt to decrease the number of filesystem operations per transaction. First, the Berkeley DB software includes support for group commit. Group commit simply means that when the information about one transaction is flushed to disk, the information for any other waiting transactions will be flushed to disk at the same time, potentially amortizing a single log write over a large number of transactions. There are additional tuning parameters which may be useful to application writers:

If you are bottlenecked on logging, the following test will help you confirm that the number of transactions per second that your application does is reasonable for the hardware on which you are running. Your test program should repeatedly perform the following operations:

The number of times that you can perform these three operations per second is a rough measure of the minimum number of transactions per second of which the hardware is capable. This test simulates the operations applied to the log file. (As a simplifying assumption in this experiment, we assume that the database files are either on a separate disk; or that they fit, with some few exceptions, into the database cache.) We do not have to directly simulate updating the log file directory information because it will normally be updated and flushed to disk as a result of flushing the log file write to disk.

Running this test program, in which we write 256 bytes for 1000 operations on reasonably standard commodity hardware (Pentium II CPU, SCSI disk), returned the following results:

% testfile -b256 -o1000
running: 1000 ops
Elapsed time: 16.641934 seconds
1000 ops:   60.09 ops per second

Note that the number of bytes being written to the log as part of each transaction can dramatically affect the transaction throughput. The test run used 256, which is a reasonable size log write. Your log writes may be different. To determine your average log write size, use the db_stat utility to display your log statistics.

As a quick sanity check, the average seek time is 9.4 msec for this particular disk, and the average latency is 4.17 msec. That results in a minimum requirement for a data transfer to the disk of 13.57 msec, or a maximum of 74 transfers per second. This is close enough to the previous 60 operations per second (which was not done on a quiescent disk) that the number is believable.

An implementation of the previous example test program for IEEE/ANSI Std 1003.1 (POSIX) standard systems is included in the Berkeley DB distribution.