Looking for TT policy advice

Discussion of chess software programming and technical issues.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
voffka
Posts: 71
Joined: Sat Jun 30, 2018 8:58 pm
Location: Ukraine
Full name: Volodymyr Shcherbyna

Re: Looking for TT policy advice

Post by voffka » Fri Oct 11, 2019 6:56 pm

Vivien,

Have you considered that the problem with long TC is not related to hash usage but rather to prunning conditions that are triggered differently in longer controls?

User avatar
xr_a_y
Posts: 768
Joined: Sat Nov 25, 2017 1:28 pm
Location: France

Re: Looking for TT policy advice

Post by xr_a_y » Fri Oct 11, 2019 7:51 pm

voffka wrote:
Fri Oct 11, 2019 6:56 pm
Vivien,

Have you considered that the problem with long TC is not related to hash usage but rather to prunning conditions that are triggered differently in longer controls?
I cannot really answer yes to that question ... It may indeed be the case. Are you thinking of a particular parameter ?

I thought TT was a good candidate because with the used replacement scheme.

User avatar
hgm
Posts: 23725
Joined: Fri Mar 10, 2006 9:06 am
Location: Amsterdam
Full name: H G Muller
Contact:

Re: Looking for TT policy advice

Post by hgm » Fri Oct 11, 2019 10:36 pm

Why would you want to store the score part? A hash move should always be searched first. An age field would be useless on an always-replace (or an under-cut) slot. And two bits would probably be large enough for it, so it could be easily packed with the bound-type flag, (whick also only needs 2 bits).

User avatar
xr_a_y
Posts: 768
Joined: Sat Nov 25, 2017 1:28 pm
Location: France

Re: Looking for TT policy advice

Post by xr_a_y » Sat Oct 12, 2019 6:05 am

hgm wrote:
Fri Oct 11, 2019 10:36 pm
Why would you want to store the score part? A hash move should always be searched first. An age field would be useless on an always-replace (or an under-cut) slot. And two bits would probably be large enough for it, so it could be easily packed with the bound-type flag, (whick also only needs 2 bits).
You are right, I was too lazy and use the main move packing also inside TT. Test in progress now with a smaller entry size.

voffka
Posts: 71
Joined: Sat Jun 30, 2018 8:58 pm
Location: Ukraine
Full name: Volodymyr Shcherbyna

Re: Looking for TT policy advice

Post by voffka » Wed Oct 16, 2019 3:29 pm

xr_a_y ,
xr_a_y wrote:
Fri Oct 11, 2019 7:51 pm
I cannot really answer yes to that question ... It may indeed be the case. Are you thinking of a particular parameter ?
I thought TT was a good candidate because with the used replacement scheme.
I think instead of spending significant time on some idea it may be worth rechecking it: if you think Topple and Igel are better in longer tc and the suspect is use of hash, re-compile them both without hash (and some other engines) and run tests in faster time control without hash (Minic, Topple, Igel, some other engines?): this will at least give you a hint if you are going a right direction.

My intiution tells me that after you run the preliminary tests you will discover that hash usage is not a culprit. When I solved that Igel problem at short time control, it was an eval problem related to stupid me doing popcount on an uninitialized variable :)

User avatar
xr_a_y
Posts: 768
Joined: Sat Nov 25, 2017 1:28 pm
Location: France

Re: Looking for TT policy advice

Post by xr_a_y » Wed Oct 16, 2019 5:19 pm

voffka wrote:
Wed Oct 16, 2019 3:29 pm
xr_a_y ,
xr_a_y wrote:
Fri Oct 11, 2019 7:51 pm
I cannot really answer yes to that question ... It may indeed be the case. Are you thinking of a particular parameter ?
I thought TT was a good candidate because with the used replacement scheme.
I think instead of spending significant time on some idea it may be worth rechecking it: if you think Topple and Igel are better in longer tc and the suspect is use of hash, re-compile them both without hash (and some other engines) and run tests in faster time control without hash (Minic, Topple, Igel, some other engines?): this will at least give you a hint if you are going a right direction.

My intiution tells me that after you run the preliminary tests you will discover that hash usage is not a culprit. When I solved that Igel problem at short time control, it was an eval problem related to stupid me doing popcount on an uninitialized variable :)
Thanks for the advice, i'll run no-TT little tourney.

User avatar
xr_a_y
Posts: 768
Joined: Sat Nov 25, 2017 1:28 pm
Location: France

Re: Looking for TT policy advice

Post by xr_a_y » Thu Oct 17, 2019 5:55 am

xr_a_y wrote:
Wed Oct 16, 2019 5:19 pm
voffka wrote:
Wed Oct 16, 2019 3:29 pm
xr_a_y ,
xr_a_y wrote:
Fri Oct 11, 2019 7:51 pm
I cannot really answer yes to that question ... It may indeed be the case. Are you thinking of a particular parameter ?
I thought TT was a good candidate because with the used replacement scheme.
I think instead of spending significant time on some idea it may be worth rechecking it: if you think Topple and Igel are better in longer tc and the suspect is use of hash, re-compile them both without hash (and some other engines) and run tests in faster time control without hash (Minic, Topple, Igel, some other engines?): this will at least give you a hint if you are going a right direction.

My intiution tells me that after you run the preliminary tests you will discover that hash usage is not a culprit. When I solved that Igel problem at short time control, it was an eval problem related to stupid me doing popcount on an uninitialized variable :)
Thanks for the advice, i'll run no-TT little tourney.
So a first test, with only 1Mb TT for everybody gave very poor results for Minic (at least -150elo very others usually at same level). So not having a big enough TT is very bad for Minic. I'll check without TT code next.

voffka
Posts: 71
Joined: Sat Jun 30, 2018 8:58 pm
Location: Ukraine
Full name: Volodymyr Shcherbyna

Re: Looking for TT policy advice

Post by voffka » Sun Oct 20, 2019 5:44 pm

150 elo diff sounds like an engine playing without any TT at all. I remember measuring the impact of tt in Igel since I completely re-written it since the fork from Greko.

User avatar
xr_a_y
Posts: 768
Joined: Sat Nov 25, 2017 1:28 pm
Location: France

Re: Looking for TT policy advice

Post by xr_a_y » Sun Oct 20, 2019 6:12 pm

voffka wrote:
Sun Oct 20, 2019 5:44 pm
150 elo diff sounds like an engine playing without any TT at all. I remember measuring the impact of tt in Igel since I completely re-written it since the fork from Greko.
Everybody had a 1Mb TT during this test. So there are two possibilities :
- either Minic TT policy is stupid for small TT
- or Minic has a serious bug that is mostly hidden by big TT

Post Reply