Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Support / Missing Hash %
- - By MarioMB [us] Date 2010-05-31 04:33
Once I bought Rybka 4 I figured I might as well use it in my Shredder interface. I also use several other engines such as Fritz 12 and Stockfish. It's obvious that Rybka is the strongest engine of the group, however when using all the other engines I get a value labeled something like "Hash XX.XX%". This is valuable because I know the status of the process. With Rybka 4 however, I don't get it. What am I doing wrong?
Parent - - By Uly (Gold) [mx] Date 2010-05-31 06:51

> This is valuable because I know the status of the process.


The information is useless because knowing how much of the hash is full doesn't tell anything about the status of the process. It just shows hash usage, but having a 100% full RAM doesn't tell you anything, since old hash entries are still overwritten.

The hash usage is misleading, that's why Rybka doesn't show it. You're not doing anything wrong.
Parent - - By GCP (***) [be] Date 2010-05-31 09:28
Well I disagree here.

Showing hash% gives you some idea if increasing the size of the hashtable makes any sense. If you're at 10% hashfull then increasing the size of the hash is unlikely to cause any visible performance difference. Whereas if 5 seconds into a 2 minute search you already hit 100%, it's quite likely you can get some improvement from increasing the hash size.
Parent - By Lukas Cimiotti (Bronze) [de] Date 2010-05-31 12:28
You are 100% right here. Vas wanted to keep things simple, so he didn't add this.
Parent - - By Banned for Life (Gold) Date 2010-05-31 17:01
Hash read hit rate is a much better metric. If increasing the hash size increases this in a meaningful way, it's a good thing to do. If it doesn't, it either won't make a difference (huge pages), or may actually hurt performance (TLB thrashing with small pages).

The biggest problem with showing hash % full is that is conveys to people that a warm hash is a bad thing. As you know, the reality is otherwise...
Parent - - By GCP (***) [be] Date 2010-05-31 19:15

>Hash read hit rate is a much better metric. If increasing the hash size increases this in a meaningful way, it's a good thing to do.


I'm not sure about this one. The problem is that a minor increase in hitrate may represent a disproportionate amount of work avoided.

It is also very variable depending on the position.

>The biggest problem with showing hash % full is that is conveys to people that a warm hash is a bad thing. As you know, the reality is otherwise...


I do not understand what you mean.
Parent - By Uly (Gold) [mx] Date 2010-05-31 20:16

> I do not understand what you mean.


It's possible that:

The user thinks 100% of the hash being used is good, leading him to use less RAM so "the status of the process" is faster. Or

The user thinks that it's good that the hash has enough space for future entries (or w/e), leading him to clear the hash which could hurt his performance.

Or I don't know what else the average user thinks about this value. In the end, the Hash % is useful to a couple of advanced users, and a confusing distraction for the rest that don't know how hash works or what this number really means.
Parent - By Banned for Life (Gold) Date 2010-05-31 22:16
The problem is that a minor increase in hitrate may represent a disproportionate amount of work avoided.

Of course you're right. I'm more used to dealing with cache issues where all cache lines are assumed to be equally important (i.e. if there's a miss, you need to go out to higher cache or memory). In the chess case, it's a lot worse overwriting a high depth entry. Maybe a depth weighted hit rate, or even better, a hit rate for each depth would resolve this problem.

It is also very variable depending on the position.

The hit rates will vary depending on the position, but it's still a good measure of how much utility you're getting from the TT in that particular position.

    >The biggest problem with showing hash % full is that is conveys to people that a warm hash is a bad thing. As you know, the reality is otherwise...

I do not understand what you mean.


When you clear the hash and start analyzing a new position, the % full is at 0. This means you will initially only have low depth entries. As time progresses, the % full will rise toward 100% and you will have a distribution of depths. The % full at 100 is clearly better for analysis than the % full at 0, while most users will see the full hash as being undesirable. In fact, one frequently asked question is whether analysis will proceed faster if hash is cleared in this case...
Parent - By Vasik Rajlich (Silver) [pl] Date 2010-06-04 14:55

> Showing hash% gives you some idea if increasing the size of the hashtable makes any sense. If you're at 10% hashfull then increasing the size of the hash is unlikely to cause any visible performance difference. Whereas if 5 seconds into a 2 minute search you already hit 100%, it's quite likely you can get some improvement from increasing the hash size.


I should probably report something, but "hashfull" is not a good figure. The question is what is being replaced. There is a big difference between 99% hashfull with lots of minimum-depth entries and 99% hashfull with lots of higher-depth entries.

Vas
Up Topic Rybka Support & Discussion / Rybka Support / Missing Hash %

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill