Fischerle 0.9.80 SE 64-bit Gauntlet for CCRL 40/40

[Graham Banks]

PGN - http://kirill-kryukov.com/chess/discuss ... p?id=39548

[RolandoFurioso]

Thanks, Graham!

[RolandoFurioso]

Hi Gabor,

thanks for this information.

I have no idea what is going wrong here. Regarding the multiple instances of Java, this is certainly no Fischerle 0.9.80 issue, but presumably a problem with the environment you are running it within. Fischerle definitely doesn't spwan additional JRE processes; most likely, your environment doesn't properly terminate the JRE process if an instance of Fischerle is terminated and a new one is started.

No idea euther why Fischerle 0.9.80 doesn't perform as expected under your particular Arena environment. Could you try using the .exe starter instead of the .bat file for starting it (... though, of course, both should worl well)? Does it play, according to your observations, much weaker than F 0.9.70? Of course, this shouldn't be the case - e. g., the recently completed CEGT testings iindicate that it plays significantly stronger than its predecessor. In fact, it should at least reach the same level of play as F 0.9.70. Thus, could you perhaps pit these two latest Fischerle versions against each other, using an identical way of installing them under Arena (.exe) and describe what happens - does F 0.9.80 really play significantly weaker . which I could hardly imagine?

Thanks & all the best
Roland

[RolandoFurioso]

PS: Could it be the case that you observed the performance issues under Arena while there were still a couple of Java (JRE) processes running on your computer? If this was the case, this could well explain the performance issues: if, due to the already running JRE instances, the physical memory is nearly exhausted, the very Fischerle instance running under Arena might run very slow because there are too many accesses to virtual (swap file) memory. So please try to run Fischerle 0.9.80 on a system with yet a sufficient amount of physicaal memory available. The Fischerle starters (.exe as well as .bat) try to assign a minimum amount of memory to the JRE process ("-Xms1400m -Xmx1400m") in order to minimize dynamic memory requests and garbage collect operations that would be detrimental to Fischerle's performance.

[sandermvdb]

I'm glad you have no problems running Fischerle. I used to have none either. However, now I seem to have some serious issue.

Under Arena, the engine's performance is way below what is expected. I tried it under WinBoard as well but then every now and then java remains in memory and very soon my memory gets filled with java instances. It is possible that there is something wrong with the UCI2WB adapter which WB currently uses (it used to be Polyglot) but I find no explanation for the performance under Arena.

Under Arena I use Roland's batch file while under WB I have this line in the WB.ini file:

"Fischerle 0.9.80 SE 64-bit" -fcp "java -Xms1400m -Xmx1400m -XX:+UseParallelGC -jar Fischerle.jar uci" -fd "E:\Sakk\UCI-engines\Fischerle\dist" -fn "Fischerle 0.9.80 SE 64-bit" -fNoOwnBookUCI -fUCI

Why do you use a batch-file under Arena? If you don't want to set any special vm parameters you can just point directly to the .jar file in Arena (I guess the only reason you want a batch-file if you want to specify a really large hashtable. The default max memory usage of java is 1/4th of the total memory which is probably more than enough)

Do you use a 64 bit JRE? You can check this by opening a command prompt and execute
java -version

It should say (latest version):
Java HotSpot(TM) 64-Bit Server VM (build 25.131-b11, mixed mode)

[RolandoFurioso]

Under Arena I use Roland's batch file while under WB I have this line in the WB.ini file:

"Fischerle 0.9.80 SE 64-bit" -fcp "java -Xms1400m -Xmx1400m -XX:+UseParallelGC -jar Fischerle.jar uci" -fd "E:\Sakk\UCI-engines\Fischerle\dist" -fn "Fischerle 0.9.80 SE 64-bit" -fNoOwnBookUCI -fUCI

Why do you use a batch-file under Arena? If you don't want to set any special vm parameters you can just point directly to the .jar file in Arena (I guess the only reason you want a batch-file if you want to specify a really large hashtable. The default max memory usage of java is 1/4th of the total memory which is probably more than enough)

Well - it definitely doesn't hurt to supply these parameters even if some of them won't have any effect in a particular environment. And there is not only the "-Xmx1400m" (= max) parameter, but as well "-Xms1400m" (min) and a further option "-XX:+UseParallelGC " that specifies the garbage collection strategy to be employed. Parallel garbage collection is prefereable since it better distributes the GC efforts, i. e., the JVM is advised not to interrupt the execution too long for GC purposes at any particular point - and this is definitely performance-critical in the context of a chess engine.

Anyway: this discusssion is pointless with respect to the issues referred to in Gabor's original posting. So please use the (equivalent) .exe or .bat starters supplied with the Fischerle distributions, since this is the official way of running this engine - and, in fact, the only way that establishes compareability of the evaluation results, since - as discussed in the previous paragraph - using a different starter might have negative effects on the performance of the engine.

[RolandoFurioso]

The Fischerle exe you provided crashes under Arena on my PC. Does it invoke the jar file in some way? I had to rename the jar files so that they are different for 0.9.70 and 0.9.80.

Yes - it refers to the jar file; so after renaming it, the .exe won't run anymore. It is completely equivalent to the original .bat script.

But it is not necessary to rename the jar file if you just keep every Fischerle distribution in its original subdirectory. In this way, the .jar archives won't get confounded, and both .exe and .bat should work equivalently.

A 40/1 20-game match between 70 and 80 ended 11-9 in favour of 70, so no significant difference. I used Sander's suggestion for that match, that is, I pointed directly to the jar files and let Arena do the rest.
As a side note, there were 112 time forfeits during these 20 games. I guess Arena messes up something here, the games are continued after the time forfeits. I can send you the PGN although I don't think it's relevant.

The time forfeits are possibly induced by the comparably fast time control. Fischerle's java process is relatively resource-intensive and might react - depending as well on the particular computer system - a bit slowly if employed via UCI and under Arena.But I have just re-checked it on my own system, and there seem to be no issues with a 40/1 time control. Perhaps you could run a few games with 40/2?

Thanks & all the best
Roland

[RolandoFurioso]

The Fischerle exe you provided crashes under Arena on my PC. Does it invoke the jar file in some way? I had to rename the jar files so that they are different for 0.9.70 and 0.9.80.

... and thanks for pointing this out! In fact, there seems to be a big risk that somebody who wants to run the latest Fischerle version merely copies the .exe starter and then ends up using the .jar file of an older distribution. The point is that the .exe files just "wrap" the respective .bat scripts but - of course! - not the whole distribution.

Who has been running the recent Fischerle 0.9.80 SE gauntlet?

[Graham Banks]

Who has been running the recent Fischerle 0.9.80 SE gauntlet?

I ran the 40/40 gauntlet.

[RolandoFurioso]

Thanks, Graham!

I'll try to employ a more transparent style in the follow-up distributions - e. g., by giving the folder containing the Fischerle.jar file a distribution-specific name so that different versions can't be accidentally confounded.