[Computer-go] pachi2

[Hideki Kato]

Jean-loup Gailly: <AANLkTikrkfwU98w8=vEHoNxLVsv-VJQg3qpfRNT+HhNn at mail.gmail.com>:
>> The bandwidth seems not a problem but how about the latency or delay?
>
>You are right, latency might be an issue. Trying udp is on my (large) list
>of things to do.
>
>> That is, usual 1 Giga bit Ethernet?  Sorry for asking again,
>> because HPC clusters' standard is 10 Giga bit or more.
>
>If the main issue is latency and not bandwidth, there should not be a big
>difference between 1Gbit and 10 Gbit Ethernet. 

Not ignorable.  Since all modern switches use store-and-forward 
architecture, the delay time caused by a switch is almost proportional 
to the (packet-size / bandwidth).  For example, the latencies of my 
GbE switch (Allied Telesis GS900-16; $200) are 2.8 and 4.5 microsecond 
with 1000 and 100 Mbps connections, respectively, for a 64-byte 
packet.  Assuming the packet-size is 9 kB (jumbo packet), the latency 
for 1000 Mbps connection is about 72 microsecond, where that of a 
low-latency 10 GbE switch (Fujitsu XG700; $25000) is estimated about 
7 microsecond.  

Hum, in 70 microsecond, almost all strong programs can do less than 
one simulation on 19x19 but some on 9x9, right? 

>Of course some other
>network architectures, such as used by Mogo, can help a lot but I am
>not using this.

Broadcasting (O(1)) is always better than binary-tree rounting of 
multiple tcp connections (O(log n)), isn't it?

Hideki
-- 
Hideki Kato <mailto:hideki_katoh at ybb.ne.jp>