Run a process on a particular subset of the available hardware.


Documentation Build Status

A command-line tool to restrict a process from using particular hardware resources. It is particularly geared towards performance-sensitive multi-core experiments, in which you may want to avoid or measure the impact of CPU features like SMT (e.g., Intel Hyper-Threading or NUMA memory. While this is possible with existing tools like hwloc-bind and numactl, it's often pretty inconvenient to get the result you want. With curb on the other hand:

$ curb --no-smt --no-numa mycommand

You can also limit the number of cores using something like:

$ curb --no-smt -n 3 mycommand

By default, curb picks cores, nodes, and anything else in a deterministic fashion (usually tending towards 0). If you want to pick randomly instead, use --randomize (or -r).

Passing arguments

Will run mycommand only on each physical core on a single NUMA node (the first one specifically). Note that if you want to pass additional arguments to the command under test, you can do so using --:

$ curb --no-smt mycommand -- --no-crash --performance better

Visualizing the core allocation

To visualize what CPUs are being used using lstopo, do:

$ curb --no-smt --no-numa lstopo -- --pid 0

That should show all active cores in green. If you don't have a window manager running, add -v at the end and look for PUs marked with (running).


Curb depends on the hwloc crate, which in turn depends on the hwloc C library.

OS Support

Curb currently only works on cfg(unix) systems, because it sets the CPU binding for the current process, and then calls exec. This should be possible to work around by instead spawning the process, getting its identifier, calling set_cpubind_for_process, and then waiting for the process to exit. However, this has a number of downsides, like the process initially being unconstrained and having to proxy things like stdin, stdout, signals, and exit codes. Suggestions for how to support non-UNIX platforms are warmly welcome!