999954-beta for Linux and Windows: http://chess.cygnitec.com/engine/ -- am trying different optimization now, and if it's positive, will note and re-upload.
...Large pages are disabled, as they are not supported using shmget() on OSX.
I have no idea about Mac OS X, but the similar code in Crafty reads:
ERROR. shmget() failed, unable to allocate a shared memory segment.
Please verify that your /proc/sys/kernel/shmmax value is large enough to allow allocating the amount of memory you are requesting.
> echo 1000000000 > /proc/sys/kernel/shmmax
will allow a segment up to one billion bytes. I recall doing this when fooling with Crafty (it worked, and w/o re-booting).
In OS X 10.3.9 and later, create a file named /etc/sysctl.conf containing these variable assignments, and reboot:
kern.sysv.shmmax=4194304
kern.sysv.shmmin=1
kern.sysv.shmmni=32
kern.sysv.shmseg=8
kern.sysv.shmall=1024
I'm not sure that this is desired value of kern.sysv.shmmax in your case, though.
It may be necessary to reboot to be able to allocate all the hugepages that is needed. This is because hugepages requires large areas of contiguous physical memory. Over time, physical memory may be mapped and allocated to pages, thus the physical memory can become fragmented. If the hugepages are allocated early in the boot process, fragmentation is unlikely to have occurred.
My guess is that this correct. Making 2GB (or even 4GB, depending on your Hugepagesize) of huge pages with "1024" (as they said) on a 4GB machine is not likely to do well if your machine has been up for awhile. You can look at /proc/buddyinfo (lists contiguous sections available) or /proc/meminfo (gives the amount of current allocated memory) to find out more (including your Hugepagesize).
Yes, it wasn't wise to simply echo it. I had to stick it in sysctl.conf and reboot. And it works fine. And I clearly see a rise in nps -- no doubt. Kind of like drinking a venti drip, bold roast.
kingliveson wrote:999954-beta for Linux and Windows: http://chess.cygnitec.com/engine/ -- am trying different optimization now, and if it's positive, will note and re-upload.
I think my post was about shared memory under Mac OS X (which is required by Rybka, at least under microwine), and my guess is that Jeremy is saying that shmget does not have a SHM_HUGETLB option with Mac OS X. I think the PowerPC has many more page-size options that just 4K/2M/1G, but I don't know how they are all made available to the user.
Some OSs, like Solaris and FreeBSD I think, automatically use huge pages in some contexts. With Linux, they are reserved by the OS, and then can be asked for by processes. This solves the fragmentation issue if you reserve them at boot time (and don't have many different programmes asking for them), though it does mean that you have a large chunk of memory that non-hugetlb processes can't utilise. The Windows way is to reserve large pages only when a VirtualAlloc call asks for it, which means that the overall memory must be sufficiently defragged at that instant for it to work.
kingliveson wrote:999954-beta for Linux and Windows: http://chess.cygnitec.com/engine/ -- am trying different optimization now, and if it's positive, will note and re-upload.
OK, using a different compiler, and testing some new flags, this build does OK against Houdini 1.03. The results must be taken with a grain of salt as error bar for 50 games is quite high. This is not the same build uploaded to the link above -- 32-bit and non-popcnt versions are unavailable. The test used no egtb, default settings for both engines except for pawnhash. Condition, games, and the build are all included in the attachment.
Further note, this popcnt build is at at least 5% slower against single-core version non-popcnt Houdini -- yet on 4 cores, it's a different story.
IvanHoe999954 has been re-uploaded. Fixed RobboTripleBases bug, and enabled additional time management function which can be seen here: http://www.open-chess.org/viewtopic.php?p=4052#p4052. The re-uploaded file has nothing to do with previous post/attachment. That has to do with using thread Parallelization.
kingliveson wrote:IvanHoe999954 has been re-uploaded. Fixed RobboTripleBases bug, and enabled additional time management function which can be seen here: http://www.open-chess.org/viewtopic.php?p=4052#p4052. The re-uploaded file has nothing to do with previous post/attachment. That has to do with using thread Parallelization.
Best Engines-5 2010
Site/ Country: AMDX6-1090T, United States
Level: Blitz 10/0
Hardware: AMD Phenom(tm) II X6 1090T Processor6x 3895 MHz with 4,096 MB Memory
Operating system: Microsoft Windows XP 64 Bit Professional Service Pack 2 (Build 3790)
robbolito wrote:Best Engines-5 2010
Site/ Country: AMDX6-1090T, United States
Level: Blitz 10/0
Hardware: AMD Phenom(tm) II X6 1090T Processor6x 3895 MHz with 4,096 MB Memory
Operating system: Microsoft Windows XP 64 Bit Professional Service Pack 2 (Build 3790)
Thanks for testing that build. I thought parallel threading which slows the nps down just a tad bit would have negative effect, but my test, your test, and that of our friend showed it to be doing good. Will investigate it a little more. It would have made more sense for me to rename the build to something that differentiates it from the others. Thanks for testing it.
I believe that Houdini is not getting its full strength on this 6-core AMD - prozessor. It is strange that it is below Rybka, whereas in other tests it is above.
