Worker NUMA binding happens too early for useful errors
wingo opened this issue · comments
Andy Wingo commented
Currently a ptree worker is bound to a NUMA node early. This means that memory will only be allocated on that NUMA node, and then existing pages are migrated to that node. However this happens before the worker has a large working set, which makes errors unlikely. Instead in the case where a worker would have a large working set it might be useful to bind the NUMA node later, in order to see errors (if any).
The alternative with a policy of MPOL_BIND
is that if the worker runs out of local memory, it gets hit by a SIGBUS. Not so nice!