Lbench

Lbench is a simple Linux multithread benchmark tool.
Various aspects of CPU and OS performance are measured for 1-64 parallel threads.
Lbench measures how performance scales with multiple threads and multiple CPU processor cores.

Lbench can make the following measurements for 1 to 64 parallel threads:

• integer-32 and float-64 arithmetic performance
• performance of common engineering functions (square root, sine, etc.)
• memory throughput for processor cache and main memory
  (block memory moves, int-32 sequential get/put, int-32 random get/put)
  
• function call/return overhead
• 2D matrix math performance
• Fibonacci benchmark (recursion method)
  
• sort millions of random character strings using 4 threads
  
• classic Whetstone benchmark for type double (64 bit floating point)
• classic Linpack benchmark for type double
• thread switch rate for do-nothing threads
  
• time required to start and complete a do-nothing process thread
• time required to start and complete a do-nothing sub-process
• time required to acquire and release a global lock
• disk throughput for serial and random I/O using various block sizes
  (direct to disk without OS caching)

Lbench has two other functions not related to benchmarking:

• Cooling performance: run multiple CPU-bound threads, monitor processor throughput and
  temperature, detect if the CPU clock is throttled down due to thermal overload.
• Memory test and burn-in: continuously fill memory with random values, read and compare.

The user guide is available from the [help] button and goes into more detail about each
benchmark and configurable parameters.

Some Conclusions based on benchmarking the author's PC (July 2025):

• For pure calculations, 4 threads have nearly 4 times the aggregate throughput of 1 thread.
• The cache memory throughput for 4 threads is nearly 4 times that of 1 thread.
• The main memory throughput for 4 threads is not much greater than 1 thread.
  
• Block memory moves are 2-20 times faster than 4-byte sequential get/put moves
  (depends on size of block relative to size of CPU cache)
  
• 4-byte sequential get/put moves are 1 - 10 times faster than 4-byte random get/put moves.
• 4-byte sequential get/put moves have about the same speed for cache and main memory.
  
• 4 threads contending for a mutex-controlled resource can swap ownership >20 million times/sec..
• The time to create a thread which does nothing but exit is about 28 microseconds.
  
• The time to create a sub-process which does nothing but exit is about 296 microseconds
  

Example Output Using 4 Threads