4 ms·
Yep in Netflix case they pack bare-metal instances with a very large amount of containers and oversubscribe them (similar to what Borg reports: hundreds of cont
by mochomocha 2y ago
Yep in Netflix case they pack bare-metal instances with a very large amount of containers and oversubscribe them (similar to what Borg reports: hundreds of containers per VM is common), so there are always more runnable threads than CPUs and your runqueues fill up.
- otterley 2y agoI'm curious as to the capacity of the bare metal hosts you operate such that you can oversubscribe CPU without exhausting memory first or forcing processes to swap (which leads to significantly worse latency than typical scheduling delays). My experience is that most machines end up being memory bound because modern software—especially Java workloads, which I know Netflix runs a lot of—can be profligate memory consumers.