Hash usage percent display (in Arena)

[Jouni]

Only Houdini and Critter are showing this value now and I noticed with "only" 3-4 Mnodes/s it goes in one minute around 95-98% with 1 GB for hash! So is this meaning too small hash size? And what is optimum hash usage % in play or in analysis?

Jouni

[Houdini]

Houdini uses 16 bytes per hash entry and has a hash fill ratio close to 50%. With a node speed of 4,000 kN/s on your computer it will fill the hash at a rate of 32 MB/s or about 2 GB/min.

If the reserved hash size is significantly smaller than this theoretical amount you'll start overwriting a lot of hash entries and the search will become less efficient. On the other hand, if you reserve too much hash memory the access will be slower and the node speed will drop.

Robert

[hgm]

My experience is that a noticeable slowdown only occurs with a loading factor of 10 or more. And with a good replacement scheme even then the search time is only a very weak function of overload. (Like the 12th root or so: to double the search time required a 4096 times smaller hash table.)

[bob]

My experience is that a noticeable slowdown only occurs with a loading factor of 10 or more. And with a good replacement scheme even then the search time is only a very weak function of overload. (Like the 12th root or so: to double the search time required a 4096 times smaller hash table.)

I have not run that test in years, but I didn't find it that forgiving. But it was not as if you go past a 100% loading that you fall off a cliff. I'll try to do this again, although Crafty might not be the best representative since it doesn't hash qsearch stuff...

[hgm]

Indeed, hashing QS might make a difference, and the loss of QS nodes from the table by very rapid overwriting when the loading factor is 10 might not hurt very much, as the static move ordering is pretty good at dentifying the best capture, and QS could at best provide a capture for hash moves. (My test was done with an engine that did not search checks in QS.)

[mhalstern]

Some quick tests on my pc show Houdini 1.5a reporting from just below 11,000 kn/s to just over 12,000/kn/s. I ran the tests using 1 gig of hash and analysing various positions for 30 seconds.

My pc has 6 gig of Ram.

Robert,

What would be the ideal hash settings for Houdini for these conditions:

1 - Long runs ( hours) of infinite analysis
2 - Short Engine Matches - G-5 to G-15
3 - Long Engine Matches G - 6 hours

Thanks

[Houdini]

Some quick tests on my pc show Houdini 1.5a reporting from just below 11,000 kn/s to just over 12,000/kn/s. I ran the tests using 1 gig of hash and analysing various positions for 30 seconds.

My pc has 6 gig of Ram.

Robert,

What would be the ideal hash settings for Houdini for these conditions:

1 - Long runs ( hours) of infinite analysis
2 - Short Engine Matches - G-5 to G-15
3 - Long Engine Matches G - 6 hours

Thanks

12,000 kN/s means the hash filling rate is about 100 MB/s.

For infinite analysis or long TC games the largest possible hash size is useful, 4 GB on your computer.

For short TC games you pick a good hash value by multiplying the average move time by the hash filling rate.
For example, if you're playing 5 minute games the average move time is about 5 seconds. On your hardware about 500 MB of hash will be written for each move, so optimum performance will be obtained with hash sizes between 256 MB and 1024 MB. As HGM noted, the exact value isn't critical.

Robert