| Author |
Thread Statistics | Show CCP posts - 20 post(s) |
ElfeGER
Black Eclipse Corp
Band of Brothers
   |
Posted - 2008.05.27 17:06:00 -
[1]
how good is windows at running multiple nodes per cpu core?
as in distribute like 50 nodes on 16 cores and let them balance in real time
I wonder if wine is able to run a solnode process as linux has several good task schedulers built in
(ages ago the beta farcry dedicated windows server was running fine on wine)
|
ElfeGER
Black Eclipse Corp
Band of Brothers
|
Posted - 2008.05.29 16:40:00 -
[2]
the issue is that there is running one process per cpu and probably bound to the cpu as well
by in creasing the processes to like 4x cores you could get have finer balancing as the os decides in realtime how the processes are distributed
so you could have running 1 process on one core in a big battle/busy system while the other core runs the other 7 processes and is still not maxed out (that's where multi processor 4x cores come in handy)
on the other side that would increase the memory, db connection usage as well as inter node jumps
|
ElfeGER
Black Eclipse Corp
Band of Brothers
|
Posted - 2008.05.29 16:43:00 -
[3]
 Originally by: Ishina Fel
For example: "Write a load balancer that can, in real time, shift a process involving SQL queries to the CPU core with the lowest load. Use only Stackless Python version (blah)"
hmmm why reinvent the wheel? the os handles that automaticly
|
ElfeGER
Black Eclipse Corp
Band of Brothers
|
Posted - 2008.05.29 18:16:00 -
[4]
Originally by: Venchoba
What if you distribute computation on the same grid on several threads. One thread for area effects (movement, smartbomb, ECM burst, interdictor bubbles) and N threads with directed/local interaction (guns, missles and anything for which you need lock).
Also, since this new cluster needs server code to be rewriten and one of big problems is I/O bandwidth, did CCP ever consider moving to mainframe? Mainframes are kings of I/O. Those things can process transactions of the credit cards in time of Christmas.
you would still need a lot of locks to get these things synced
but talking of io bandwidth
what bandwidth/processing do you need when one person activates 8 guns in a fleet of 400 ppl?
8 * 400 = 3200 that player x activated each gun on target y?
now let's say 95% have turret effects disabled and do the math again
8 * 20 = 160
now let's say there is a magic way to do player x activated all his guns on target y
1 * 20 = 20
now let's do these number with actually 400 ppl doing something
so about
1280000
64000
8000
then add movement, drones, reloading, ship destructions, targeting, gun deactivations and new targets 
|
| |
|