On 1/21/20 2:50 PM, Peter Zijlstra wrote: > On Tue, Jan 21, 2020 at 02:29:49PM +0100, Peter Zijlstra wrote: >> On Mon, Dec 30, 2019 at 02:40:41PM -0500, Alex Kogan wrote: >> >>> +/* >>> + * Controls the threshold for the number of intra-node lock hand-offs before >>> + * the NUMA-aware variant of spinlock is forced to be passed to a thread on >>> + * another NUMA node. By default, the chosen value provides reasonable >>> + * long-term fairness without sacrificing performance compared to a lock >>> + * that does not have any fairness guarantees. The default setting can >>> + * be changed with the "numa_spinlock_threshold" boot option. >>> + */ >>> +int intra_node_handoff_threshold __ro_after_init = 1 << 16; >> There is a distinct lack of quantitative data to back up that >> 'reasonable' claim there. >> >> Where is the table of inter-node latencies observed for the various >> values tested, and on what criteria is this number deemed reasonable? >> >> To me, 64k lock hold times seems like a giant number, entirely outside >> of reasonable. > Daniel, IIRC you just did a paper on constructing worst case latencies > from measuring pieces. Do you have data on average lock hold times? > I am still writing the paper, but I do not have the (avg) lock times. It is it is in the TODO list, though! -- Daniel
