DESCRIPTIONlmbench is a series of micro benchmarks intended to measure basic operating system and hardware system metrics. The benchmarks fall into three general classes: bandwidth, latency, and ``other''.
Most of the lmbench benchmarks use a standard timing harness described in timing(3) and have a few standard options: parallelism, warmup, and repetitions. Parallelism specifies the number of benchmark processes to run in parallel. This is primarily useful when measuring the performance of SMP or distributed computers and can be used to evaluate the system's performance scalability. Warmup is the number of minimum number of microseconds the benchmark should execute the benchmarked capability before it begins measuring performance. Again this is primarily useful for SMP or distributed systems and it is intended to give the process scheduler time to "settle" and migrate processes to other processors. By measuring performance over various warmup periods, users may evaulate the scheduler's responsiveness. Repetitions is the number of measurements that the benchmark should take. This allows lmbench to provide greater or lesser statistical strength to the results it reports. The default number of repetitions is 11.
BANDWIDTH MEASUREMENTSData movement is fundamental to the performance on most computer systems. The bandwidth measurements are intended to show how the system can move data. The results of the bandwidth metrics can be compared but care must be taken to understand what it is that is being compared. The bandwidth benchmarks can be reduced to two main components: operating system overhead and memory speeds. The bandwidth benchmarks report their results as megabytes moved per second but please note that the data moved is not necessarily the same as the memory bandwidth used to move the data. Consult the individual man pages for more information.
Each of the bandwidth benchmarks is listed below with a brief overview of the intent of the benchmark.
- reading and summing of a file via the read(2) interface.
- memory copy.
- memory reading and summing.
- memory writing.
- reading and summing of a file via the memory mapping mmap(2) interface.
- reading of data via a pipe.
- reading of data via a TCP/IP socket.
- reading data from a UNIX socket.
LATENCY MEASUREMENTSControl messages are also fundamental to the performance on most computer systems. The latency measurements are intended to show how fast a system can be told to do some operation. The results of the latency metrics can be compared to each other for the most part. In particular, the pipe, rpc, tcp, and udp transactions are all identical benchmarks carried out over different system abstractions.
Latency numbers here should mostly be in microseconds per operation.
- the time it takes to establish a TCP/IP connection.
- context switching; the number and size of processes is varied.
- fcntl file locking.
- ``hot potato'' transaction through a UNIX FIFO.
- creating and deleting small files.
- the time it takes to fault in a page from a file.
- memory read latency (accurate to the ~2-5 nanosecond range, reported in nanoseconds).
- time to set up a memory mapping.
- basic processor operations, such as integer XOR, ADD, SUB, MUL, DIV, and MOD, and float ADD, MUL, DIV, and double ADD, MUL, DIV.
- ``hot potato'' transaction through a Unix pipe.
- process creation times (various sorts).
- ``hot potato'' transaction through Sun RPC over UDP or TCP.
- select latency
- signal installation and catch latencies. Also protection fault signal latency.
- non trivial entry into the system.
- ``hot potato'' transaction through TCP.
- ``hot potato'' transaction through UDP.
- ``hot potato'' transaction through UNIX sockets.
- the time it takes to establish a UNIX socket connection.
- processor cycle time
- TLB size and TLB miss latency
- cache line size (in bytes)
- cache statistics, such as line size, cache sizes, memory parallelism.
- John McCalpin's stream benchmark
- memory subsystem parallelism. How many requests can the memory subsystem service in parallel, which may depend on the location of the data in the memory hierarchy.
- basic processor operation parallelism.
ACKNOWLEDGEMENTFunding for the development of these tools was provided by Sun Microsystems Computer Corporation.
A large number of people have contributed to the testing and development of lmbench.
COPYINGThe benchmarking code is distributed under the GPL with additional restrictions, see the COPYING file.
AUTHORCarl Staelin and Larry McVoy
Comments, suggestions, and bug reports are always welcome.