Re: Congrats the Komodo Team
Posted: Sat Aug 17, 2019 3:17 pm
It's a benefit of the new UPI interconnect, however the actual latency seems to be higher on average than the earlier QPI version, at least for TT-style access patterns.zullil wrote: ↑Sat Aug 17, 2019 12:07 pmThanks. So all other nodes are "equidistant" from any one fixed node. So even if a given search thread wanders off its preferred node, the penalty is never too severe.mjlef wrote: ↑Sat Aug 17, 2019 3:25 amhere they are:zullil wrote: ↑Fri Aug 16, 2019 12:09 pmOut of curiosity, what are the "node distances" on that machine? Under Linux, that table is provided by runningmjlef wrote: ↑Fri Aug 16, 2019 6:24 am I think I need to do some more programming to take advantage of an 8 NUMA node machine. All those cores trying to access memory requires more speical memory management, so we will work on it. Komodo was getting about 70-90 million nps for much of the positions, but faster in endgames. With some work it can probably do better, but each processor doubling gives less and less elo and at some point, more is less.Code: Select all
numactl --hardware
node distances:
node 0 1 2 3 4 5 6 7
0: 10 20 20 20 20 20 20 20
1: 20 10 20 20 20 20 20 20
2: 20 20 10 20 20 20 20 20
3: 20 20 20 10 20 20 20 20
4: 20 20 20 20 10 20 20 20
5: 20 20 20 20 20 10 20 20
6: 20 20 20 20 20 20 10 20
7: 20 20 20 20 20 20 20 10