Knps slowdown

[cc2150dx]

Intel i7 3770 quad core
8 gigs ram
GUI..Arena 3.5
both engines 1-cpu

I have been doing 60 second runs (from start position) of Komodo 9.1 and Stockfish 6 by doubling the ram each time and been noticing Knps slowdown.

Komodo:

1024 MB..1574 Knps
2048.......1534
4096.......1497

Stockfish:

1024 MB..1767 Knps
2048.......1684
4096.......1639

Does this seem right? Can I even use those large sizes of ram without performance loss?


Thanks,


Jason

[bob]

Intel i7 3770 quad core
8 gigs ram
GUI..Arena 3.5
both engines 1-cpu

I have been doing 60 second runs (from start position) of Komodo 9.1 and Stockfish 6 by doubling the ram each time and been noticing Knps slowdown.

Komodo:

1024 MB..1574 Knps
2048.......1534
4096.......1497

Stockfish:

1024 MB..1767 Knps
2048.......1684
4096.......1639

Does this seem right? Can I even use those large sizes of ram without performance loss?


Thanks,


Jason

Yes. One problem is TLB. As you increase the number of pages, the TLB gets swamped. Going to 1gb pages will help this markedly. However, Knps is not the end of the story. You will likely notice that even though Knps drops, time to depth improves. The ideal hash size is the one that minimizes time to depth for a given time limit per move.

[cc2150dx]

Intel i7 3770 quad core
8 gigs ram
GUI..Arena 3.5
both engines 1-cpu

I have been doing 60 second runs (from start position) of Komodo 9.1 and Stockfish 6 by doubling the ram each time and been noticing Knps slowdown.

Komodo:

1024 MB..1574 Knps
2048.......1534
4096.......1497

Stockfish:

1024 MB..1767 Knps
2048.......1684
4096.......1639

Does this seem right? Can I even use those large sizes of ram without performance loss?


Thanks,


Jason

Yes. One problem is TLB. As you increase the number of pages, the TLB gets swamped. Going to 1gb pages will help this markedly. However, Knps is not the end of the story. You will likely notice that even though Knps drops, time to depth improves. The ideal hash size is the one that minimizes time to depth for a given time limit per move.

Hello Robert (Bob),

Thanks for the detailed information. I'm a little rusty on my computer terms so I don't know what TLB means.


Jason

[bob]

Intel i7 3770 quad core
8 gigs ram
GUI..Arena 3.5
both engines 1-cpu

I have been doing 60 second runs (from start position) of Komodo 9.1 and Stockfish 6 by doubling the ram each time and been noticing Knps slowdown.

Komodo:

1024 MB..1574 Knps
2048.......1534
4096.......1497

Stockfish:

1024 MB..1767 Knps
2048.......1684
4096.......1639

Does this seem right? Can I even use those large sizes of ram without performance loss?


Thanks,


Jason

Yes. One problem is TLB. As you increase the number of pages, the TLB gets swamped. Going to 1gb pages will help this markedly. However, Knps is not the end of the story. You will likely notice that even though Knps drops, time to depth improves. The ideal hash size is the one that minimizes time to depth for a given time limit per move.

Hello Robert (Bob),

Thanks for the detailed information. I'm a little rusty on my computer terms so I don't know what TLB means.


Jason

Translation look-aside buffer. This is used to speed up virtual address to real address translations. With no TLB, these add four(!!) memory accesses to a normal memory fetch, four to translate the virtual address to a real/physical address, and then the actual memory access to read/write the data. The TLB eliminates those extra memory accesses speeding things up. Note that the 4 accesses will mainly be handled by cache, but that is not free. The TLB is.

[Jesse Gersenson]

Time to depth can be checked from the command line:

# start komodo
komodo.exe

# check settings
uci

# tell it to search to depth 24
go depth 24

# set Hash to 128 mb
setoption name Hash value 128

# quit
quit

[gordonr]

Time to depth can be checked from the command line:

# start komodo
komodo.exe

# check settings
uci

# tell it to search to depth 24
go depth 24

# set Hash to 128 mb
setoption name Hash value 128

# quit
quit

In my experience, I've found the process to be more involved in order to get a reliable value. My initial approach was unreliable...

http://talkchess.com/forum/viewtopic.ph ... highlight=

I now use a larger range of positions and use the median after multiple runs (automated - not a manual task).

[zullil]

Time to depth can be checked from the command line:

# start komodo
komodo.exe

# check settings
uci

# tell it to search to depth 24
go depth 24

# set Hash to 128 mb
setoption name Hash value 128

# quit
quit

Probably want to set the hash size before the "go depth 24".