Known Issues

As already discussed throughout the documentation, there are some known issues in the benchmark. The important ones are

  • It is imperative to use --fork-writer option to achieve high data throughput with multiple writers. This limitation comes from the global lock of the HDF5 libraries.

  • The option --fork-writer does not work on Mac as dispatch_semaphore library used for the implementing locks on writers supports only threads. Typically, the writer processes hangs waiting for the lock to release. One possible solution to this is to use named semaphores from POSIX (MacOS supports only named Semaphores). But using named Semaphores gives us another limitation as typically the number of opened file descriptors are limited. To get around this issue we can increase the ulimit -n (till the value the kernel accepts) and/or decrease visibility queue size for the benchmark --visibility-queue.

  • Often when running on MacOS, OpenMPI that is build with homebrew might throw some silent errors concerning vader BTL component. They might not be critical but to get rid of them include only tcp and self BTL components using --mca btl tcp,self argument.