CCRL Discussion Board

I thought I would share these results here as they are very interesting...

Hardware.

Q6600 at 3.2Ghz - (9 x 356 @ 44412)
Gigabyte P965 DS3

HD Seagate 7200.10 16mb 320GB
and
Flash EMTEC Intuix S520 8 GB USB 2.0

Each engine was allocated 2GB Hash, or the maximum it could use (Naum 1GB).
5 Men TBs on either HD or Flash Memory
TB Cache 32MB

The machine was restarted between each test and any startup activity was allowed to finish before the test was started, the machine also had minimum background tasks e.g. no AV etc (standalone machine no network connection).

Attached are some Graphs + raw data

The position

8/3k4/p6r/5R2/PP5p/7K/8/8 w - - 0 1

was chosen as it causes a lot of TB access.

It should be noted that this test is a worst case scenario as in an actual game the TBs will be cached in any spare physical ram - and this explains the increase in nodes/sec as the test progresses.

Shaun

P.S. I will Update as I run more tests and I have a seconds Flash drive for comparison arriving next week...

It seems that the improvement is quite substantial.
However it may be that this position is too oriented towards tablebases access.
Do you plan to test with some endgame suite ?

megamau wrote:It seems that the improvement is quite substantial.
However it may be that this position is too oriented towards tablebases access.
Do you plan to test with some endgame suite ?

Hi,

yes I agree. In fact the position was chosen because it caused a lot of TB access - I wanted to see the difference between hard drive and Flash memory performance. How much TBs effect performance and therefore how much difference fast or slow TB access makes is a different question

Shaun

A couple of the graphs above have been updated with additional test results.

Two different flash drives have been tested plus splitting the TB's across 2 flash drives to see how much that helps.

Key: I have also added some additional graphs showing time to ply, I think this is particularly intersting as it shows the effect of TBs on depth reached. Looking at the Hiarcs and Zappa results it would seem to suggest that having the TB's on a single HD may actually hurt performance compared to not having the TBs. Where as flash shows improvement, I assume this is due to processors having improved dramatically more than hard drives. A fast raid array is likely to show a different story...

I doubt the actual effect in engine matches would be more than a few elo but I am probably going to run a test or two just to check.

Shaun