| Pages: 1 2 [3] 4 5 6 :: one page |
|
|
| Author |
Thread Statistics | Show CCP posts - 13 post(s) |
Khlitouris RegusII
   |
Posted - 2009.02.16 21:11:00 -
[61]
So are you on infiniband now or is that still to come and if so can you upgrade the new ramsan 500 to infiniband or do you envisage purchasing another infiniband capable ramsan 500 in the future?
|
VaderDSL
THE GRAIL SEEKER5
|
Posted - 2009.02.16 21:30:00 -
[62]
So how does it work atm regarding the single CPU, you have one node aka solar system running on one processor? Is it physically impossible to have one node, aka solar system split across multiple cores/processors?
|
Leonid Andreyev
|
Posted - 2009.02.16 21:30:00 -
[63]
 Originally by: CCP Mindstar
Originally by: Altomat
How much performance improvement do you except
for the Nehalem Blades over the Wolfdales?
If they perform nicely, will the remaining AMD blades
upgraded to Nehalem ones?
Altomat
That is actually a big unknown for us at this point. We have had various experts in the field claiming anything between a 5% and 100%+ performance increase.
Originally by: Lord Matrix
Clock for clock, Nehalem CPUs are about 5-10% faster than previous generation. The improvement is greater in multithreading applications.
About what I have heard (and seen) at work.
Originally by: CCP Mindstar
One of the things about the EVE server is that it is essentially a single-threaded app at this point in time. The Wolfdales are very good for EVE as they are simply very high clock speeds. The newer processors coming out from AMD and Intel tend to focus more on multiple cores than on high clock speed. That is great for multi-threaded applications, but not so much for EVE.
The Nehalems do have a few interesting tricks up their sleeve - like turning off all but 1 core and then overclocking the one that is still running - which may give us quite a boost. Until we get some real world tests in, we don't really have any solid idea.
I wonder whether the overclocking of Core 0 and turning off other cores would be better than moving the EVE server apps to multi-threading?
|
BenjaminBarker
|
Posted - 2009.02.16 21:36:00 -
[64]
Originally by: Leonid Andreyev
...
I wonder whether the overclocking of Core 0 and turning off other cores would be better than moving the EVE server apps to multi-threading?
I love this blog post: Hardware is Cheap, Programmers are Expensive |
Simmith
Macabre Votum
Morsus Mihi
|
Posted - 2009.02.16 22:09:00 -
[65]
Originally by: DrAtomic
Could your pretty pretty pretty please take some pix to share with us from the racks (both front and behind) next time you guys are in London? Or pix or stfu as they say! hehehe ;)
Who knows it might be photography forbidden in the server halls at the hosting company. But if not.. pics!
LP Store Database | ME Evaluator |
Kweel Nakashyn
Minmatar
Kernel of War
Tau Ceti Federation
|
Posted - 2009.02.16 22:21:00 -
[66]
Originally by: CCP Mindstar
Originally by: N'olive
Quote:
In total, we now have 32 of these ultra powerful blades in the cluster - 26 running as SOL blades, and 8 running as Proxy blades.
26+8 = 34, not 32 ;)
I'm good at servers not addition 
oh my god, this is actually the worst apologize I saw this day.
And I'm reading Discworld, feat Rincewind...
16 - 12 ?
:p
Fetchez la vache !
moar(tm) > soon(tm) \o/
|
Arimathea Anthalas
Game-Over
|
Posted - 2009.02.16 23:13:00 -
[67]
Quote:
Our network engineers are busily finalizing the plans on some Cisco 7606-S routers which will be placed as the primary routing points within our game server network, and they are also revising our Internet peering strategy to expose Tranquility to our customers in the least number of hops around the globe. Lower ping times mean faster pew pew!
I'm not convinced this is a good idea. The 7606 might have some platform challenges for you. You might want to consider some alternative options for switches (Cisco Nexus, other vendors) and for the routing side, I think a 12K, CRS-1, or a modern JunOS router is a better choice. Deploying an IOS-based platform might be a mixed bag.
|
Star Nove
Minmatar
Blueprint Haus
Shadow of xXDEATHXx
|
Posted - 2009.02.16 23:36:00 -
[68]
Edited by: Star Nove on 16/02/2009 23:36:19
I'd love to know where you host this lot? Telecity/Redbus, Telehouse, Global crossing, level3, saavis etc? I worked in the industry for a long long time and have been to most of these facilities..
Awesome amount of hardware speed but I wonder what sort of connection to the net you have..? Have you thought of getting PI space and multihoming?
|
StarStryder
Ultimate Fighting Dawgs
|
Posted - 2009.02.16 23:50:00 -
[69]
Originally by: CCP Mindstar
Originally by: Altomat
How much performance improvement do you except
for the Nehalem Blades over the Wolfdales?
If they perform nicely, will the remaining AMD blades
upgraded to Nehalem ones?
Altomat
That is actually a big unknown for us at this point. We have had various experts in the field claiming anything between a 5% and 100%+ performance increase.
One of the things about the EVE server is that it is essentially a single-threaded app at this point in time. The Wolfdales are very good for EVE as they are simply very high clock speeds. The newer processors coming out from AMD and Intel tend to focus more on multiple cores than on high clock speed. That is great for multi-threaded applications, but not so much for EVE.
The Nehalems do have a few interesting tricks up their sleeve - like turning off all but 1 core and then overclocking the one that is still running - which may give us quite a boost. Until we get some real world tests in, we don't really have any solid idea.
The current AMDs within the cluster are quite fine for their task, but they will eventually need replacing at some point. When we need to replace them, we will have really good knowledge through testing of the best processors to put in.
This may be a naive thought, but can't you just run more than one copy of EVE on each blade and let each one have it's own core? Or do you run into network or memory bottlenecks doing that?
|
Lethos Aranis
THE BLACK RAGE
FOUNDATI0N
|
Posted - 2009.02.17 00:56:00 -
[70]
Edited by: Lethos Aranis on 17/02/2009 00:56:51
Originally by: CCP Mindstar
Originally by: Altomat
How much performance improvement do you except
for the Nehalem Blades over the Wolfdales?
If they perform nicely, will the remaining AMD blades
upgraded to Nehalem ones?
Altomat
That is actually a big unknown for us at this point. We have had various experts in the field claiming anything between a 5% and 100%+ performance increase.
One of the things about the EVE server is that it is essentially a single-threaded app at this point in time. The Wolfdales are very good for EVE as they are simply very high clock speeds. The newer processors coming out from AMD and Intel tend to focus more on multiple cores than on high clock speed. That is great for multi-threaded applications, but not so much for EVE.
The Nehalems do have a few interesting tricks up their sleeve - like turning off all but 1 core and then overclocking the one that is still running - which may give us quite a boost. Until we get some real world tests in, we don't really have any solid idea.
The current AMDs within the cluster are quite fine for their task, but they will eventually need replacing at some point. When we need to replace them, we will have really good knowledge through testing of the best processors to put in.
And that's the beauty of the Nehalems, the fact that you are future proofing with them if and when you eventually make Eve multi-threaded. I'd hope sooner than later because they more you wait the harder it will be and the future of computers is obviously multi-threading.
The bit about turning off all cores bar the main one on the Nehalems is fantastic though. I was reading about it the day the info was launched and was quite happy with the way it's been designed.
|
|
|
Re'taka
Minmatar
Republic University
|
Posted - 2009.02.17 01:08:00 -
[71]
Originally by: StarStryder
Originally by: CCP Mindstar
Originally by: Altomat
How much performance improvement do you except
for the Nehalem Blades over the Wolfdales?
If they perform nicely, will the remaining AMD blades
upgraded to Nehalem ones?
Altomat
That is actually a big unknown for us at this point. We have had various experts in the field claiming anything between a 5% and 100%+ performance increase.
One of the things about the EVE server is that it is essentially a single-threaded app at this point in time. The Wolfdales are very good for EVE as they are simply very high clock speeds. The newer processors coming out from AMD and Intel tend to focus more on multiple cores than on high clock speed. That is great for multi-threaded applications, but not so much for EVE.
The Nehalems do have a few interesting tricks up their sleeve - like turning off all but 1 core and then overclocking the one that is still running - which may give us quite a boost. Until we get some real world tests in, we don't really have any solid idea.
The current AMDs within the cluster are quite fine for their task, but they will eventually need replacing at some point. When we need to replace them, we will have really good knowledge through testing of the best processors to put in.
This may be a naive thought, but can't you just run more than one copy of EVE on each blade and let each one have it's own core? Or do you run into network or memory bottlenecks doing that?
As far as nodes gos, each core does run a node, on its own.
|
Chainsaw Plankton
IDLE GUNS
IDLE EMPIRE
|
Posted - 2009.02.17 02:24:00 -
[72]
Originally by: CCP Mindstar
One of the things about the EVE server is that it is essentially a single-threaded app at this point in time.
Oh really 
|
Darth Sith
Genbuku.
Daisho Syndicate
|
Posted - 2009.02.17 02:33:00 -
[73]
Originally by: LaVista Vista
Originally by: Zex Maxwell
Quote:
meaning the entire EVE database will be run from solid-state storage! Awesome!
NO WAY!
after reading the blog: I'm not sure if i heard this right back in school. but.
I know solid-state drives are faster and all, but isn't there a higher chance for solid-state drives to fail then normal drives?
Lets see:
A normal hard-drive has has moving parts.
A solid-state drive doesn't.
A solid-state drive is basically a whole bunch of RAM.
In the real world we are seeing about 1/10 th the failure rate on SSD's vs physical disks. Ware-leveling and sector sparing have pretty much eliminated the lifespan issues of older SSD's and the limited number of times you can write to a given area on the device.
|
Darth Sith
Genbuku.
Daisho Syndicate
|
Posted - 2009.02.17 02:40:00 -
[74]
Originally by: BenjaminBarker
Originally by: Leonid Andreyev
...
I wonder whether the overclocking of Core 0 and turning off other cores would be better than moving the EVE server apps to multi-threading?
I love this blog post: Hardware is Cheap, Programmers are Expensive
That is the beuty of the Nehalems. In cases where there is a multithreaded workload placed on the chip, it will chew away as a multi core proc (each is 4 cores + 4 hyperthreads). If the schedulers determine that cores are being unused, it will shut down cores and clock up the remaining cores while maintaining the power envelope of the package (IE: 90W package is capped at 90w ). So you can ultimately end up in a situation where a single core + it's associated logical processor can be clocked up 3 speed bins while the remaining cores are powered off. So a 2.96 Ghz can end up running as a 3.33 single core + hyperthread for those single threaded applications.
And for anyone thinking that the 3.33 Wolfdale is within 5-10% of a 3.33 Nehalem EP ... prepare for a wakeup call .. this new proc screams and we are seeing some benchmarks at > 2x in some workloads in our lab.
Then you can get into a whole conversation on how the new 2 socket nehalems will have up to 4 x the memory bandwidth of the existing 775 sockets plus they use DDR3 whith a much lower loaded latency :P
But what do I know ... 
|
Jasonwilliams
|
Posted - 2009.02.17 02:42:00 -
[75]
Edited by: Jasonwilliams on 17/02/2009 02:45:37
Originally by: CCP Mindstar
That is actually a big unknown for us at this point. We have had various experts in the field claiming anything between a 5% and 100%+ performance increase.
One of the things about the EVE server is that it is essentially a single-threaded app at this point in time. The Wolfdales are very good for EVE as they are simply very high clock speeds. The newer processors coming out from AMD and Intel tend to focus more on multiple cores than on high clock speed. That is great for multi-threaded applications, but not so much for EVE.
The Nehalems do have a few interesting tricks up their sleeve - like turning off all but 1 core and then overclocking the one that is still running - which may give us quite a boost. Until we get some real world tests in, we don't really have any solid idea.
The current AMDs within the cluster are quite fine for their task, but they will eventually need replacing at some point. When we need to replace them, we will have really good knowledge through testing of the best processors to put in.
So why isn't Eve multi-threaded yet? I know it takes lots of man hours to do it, but wouldn't it help increase performance and productivity?
|
Hi Lo
Faulty Solutions
|
Posted - 2009.02.17 03:14:00 -
[76]
I have noticed the game has been a lot snappier!
Karl Kopalnia > omg Hi Lo is the name I use for all my gaming characters for 15 years--you're the one that took it :*( |
Elaron
Jericho Fraction
The Star Fraction
|
Posted - 2009.02.17 05:02:00 -
[77]
Originally by: CCP Mindstar
That is actually a big unknown for us at this point. We have had various experts in the field claiming anything between a 5% and 100%+ performance increase.
One of the things about the EVE server is that it is essentially a single-threaded app at this point in time. The Wolfdales are very good for EVE as they are simply very high clock speeds. The newer processors coming out from AMD and Intel tend to focus more on multiple cores than on high clock speed. That is great for multi-threaded applications, but not so much for EVE.
I take it the server software is profiled, so you know where it is bound most particularly: raw CPU power, processor cache, IO or memory bandwidth? The answer to that should give you a good idea how much of an advantage a Core i7 Xeon will give over Wolfdale.
|
Anneke Goulet
CUTLASS CORPORATION
|
Posted - 2009.02.17 05:43:00 -
[78]
Originally by: CCP Claw
He wasn't alone <3
So it's like this but as a threesome? ctrl-alt-del from 2006
|
Cadde
Gallente
Gene Works
AKA-AHN KINGDOM
|
Posted - 2009.02.17 07:36:00 -
[79]
Sound sweet!
I would also like to know why the server application is still a single threaded beast? To my knowledge multithreading it would reduce latencies even further and would possibly eliminate any sudden mega burst freezes?
also...
MOAR PICTURES
---------------
Opinions? Yes they belong to me, not my corp! |
Maria Kalista
|
Posted - 2009.02.17 10:24:00 -
[80]
Just don't forget to make a back-up will ya?
Fiddeling around with all that windoze $oftware makes me nervous.
And, oh yes: pix or it never happened. 
Originally by: AkRoYeR
...the beauty of EvE. You have to live on the edge all the time. If you don't stay frosty, you will die!
Best game ever!
|
|
|
Vaerah Vahrokha
Minmatar
Dark-Rising
|
Posted - 2009.02.17 10:36:00 -
[81]
Quote:
I would also like to know why the server application is still a single threaded beast? To my knowledge multithreading it would reduce latencies even further and would possibly eliminate any sudden mega burst freezes?
Because porting a single thread architecture / application into multi-thread is the nearest definition of brutal, slow and expensive torture I can find for the modern days.
Just imagine an unforeseen race condition and deadlock in the middle of a 400 v 400 fight...
|
Murtough Galaktikus
Aliastra
|
Posted - 2009.02.17 10:52:00 -
[82]
Mozo! 
|
Yayoi Tan
FACTION Inc.
|
Posted - 2009.02.17 11:17:00 -
[83]
In the terms of databases, may I ask how many SQL queries per second/hour EVE takes :) I would believe this is an astronomic number LOL but a bit curious.
|
Flinchey
Amarr
|
Posted - 2009.02.17 11:18:00 -
[84]
Originally by: LaVista Vista
Originally by: Zex Maxwell
Quote:
meaning the entire EVE database will be run from solid-state storage! Awesome!
NO WAY!
after reading the blog: I'm not sure if i heard this right back in school. but.
I know solid-state drives are faster and all, but isn't there a higher chance for solid-state drives to fail then normal drives?
Lets see:
A normal hard-drive has has moving parts.
A solid-state drive doesn't.
A solid-state drive is basically a whole bunch of RAM.
no an SSD is a whole bunch of flash memory
Originally by: Kaahles
Originally by: LaVista Vista
Originally by: Zex Maxwell
Quote:
meaning the entire EVE database will be run from solid-state storage! Awesome!
NO WAY!
after reading the blog: I'm not sure if i heard this right back in school. but.
I know solid-state drives are faster and all, but isn't there a higher chance for solid-state drives to fail then normal drives?
Lets see:
A normal hard-drive has has moving parts.
A solid-state drive doesn't.
A solid-state drive is basically a whole bunch of RAM.
This!
Only problem is if power cuts then they are empty afaik but you usually got some UPS for that kind of hardware and if I'm informed corretly they still backup on conventional hard drives all the time.
no, the RAMSAN type systems have a RAM cache (64gb im guessing CCP would get)
and 2tb of flash SSD for the actual database storage etc...
CCP (well, no company in general) would store their game server purely on RAM to only risk it being completely wiped, then requiring an entirely 'nother 6 hour downtime to re-flash.
and to anyone saying SSD's life is low blah blah... they're about par with the MTBF of a conventional HDD, with MUCH greater performance (IE, an SSD RAID0 will max out a SATA 3.0 interface.)
|
Hugh Ruka
Exploratio et Industria Morispatia
|
Posted - 2009.02.17 11:36:00 -
[85]
Originally by: LaVista Vista
Originally by: Kaahles
This!
Only problem is if power cuts then they are empty afaik but you usually got some UPS for that kind of hardware and if I'm informed corretly they still backup on conventional hard drives all the time.
I'm sure CCP has some huge-*** UPS'
Any disk array worth it's price has:
1. battery backed cache
2. cache flush mechanism that flushes any dirty pages to backend while it still can
Also write cache is always write through, never write back. This means that in case of controller failure or cache failure, you can change the controller and plug in the old battery backed cache, or you throw out the cache because it is anyway flushed to disk and you lose minimal data (or none if the cache was proper ECC etc, the controller stops using it on first problems).
--- SIG ---
CSM: your support is needed ! |
Hugh Ruka
Exploratio et Industria Morispatia
|
Posted - 2009.02.17 11:56:00 -
[86]
Originally by: Rude Bwoy
instead of running a windows 'node' on each blade, incurring Windows license fees / all that patching, why don't you use two whopping big UNIX boxes ?
actualy their setup as to the hardware part is very good. they distribute the load and in case a particular node has problems, they can upgrade a single node, if more start having problems, they change all nodes to have headroom.
It also looks quite fault tolerant, but since it is MS technology, I guess they cannot disconect one node from the cluster, repair it and then start again and join to cluster, with only a few solar systems being "out of order".
--- SIG ---
CSM: your support is needed ! |
Lost Hamster
Serenity and Hungarian Operational Team
|
Posted - 2009.02.17 11:59:00 -
[87]
I guess it look's like this:
Just in multiple rows. :)
|
Lord Fitz
Project Amargosa
|
Posted - 2009.02.17 12:11:00 -
[88]
Originally by: Jasonwilliams
So why isn't Eve multi-threaded yet? I know it takes lots of man hours to do it, but wouldn't it help increase performance and productivity?
Remember the POS exploit ? The one caused by excessive code optimisation ? Well those are far easier to track down than errors in multi-threaded code. Worse is that the errors may not be obvious until the effect is huge. It's a good goal, but needs to be worked on slowly and carefully. Plenty of other applications service a limited number of users, or work fine if a process falls over and has to be restarted, the nature of Eve (large number of users, persistent, users interacting with each other, real time interactions) means that it is much harder than most software.
Plenty of other applications are either very simple (banking) or are fine if they fall over and have to be restarted (number crunching) or can deal with cached data (web servers). There aren't very many (if any) applications out there quite like Eve.
Plenty of people blame downtimes and migrations on the operating system or database, but they haven't a clue, hundreds of thousands of people use the same operating system and database and manage with no downtime, but their tasks are always infinitely simpler, even if their databases can be larger. Also Eve was explicitly written to require downtimes for certain functions.
It makes no sense to multi-thread Eve until processors stop getting faster, and we only see improvements through parallelism, which unfortunately seems to be creeping up on us very quickly.
Obviously there are multiple threads to eve for each solar system / market etc etc. But within that node it's the one thread, sending you to another solar system on another thread, is where it gets expensive. 'session change already in progress' anyone ?
|
Ephidaurus
|
Posted - 2009.02.17 12:17:00 -
[89]
With all that I/O issues I wonder why CCP is not working towards a Mainframe setup. To me BladeCenter seems to be the wrong way to go for various reasons like cooling and such. Mainframe by far are more integrated and all the "cool" blades could be consolidated on a single Mainframe. And this machines are mainly developed for high I/O rates rather then only computation power. That should boost the I/O much better.
|
Lord Fitz
Project Amargosa
|
Posted - 2009.02.17 12:22:00 -
[90]
Originally by: Hugh Ruka
Originally by: Rude Bwoy
instead of running a windows 'node' on each blade, incurring Windows license fees / all that patching, why don't you use two whopping big UNIX boxes ?
actualy their setup as to the hardware part is very good. they distribute the load and in case a particular node has problems, they can upgrade a single node, if more start having problems, they change all nodes to have headroom.
It also looks quite fault tolerant, but since it is MS technology, I guess they cannot disconect one node from the cluster, repair it and then start again and join to cluster, with only a few solar systems being "out of order".
This is incorrect. It has nothing to do with MS technology, I'm amazed you can tie your own shoelaces with knowledge of the IT industry like that. The eve server written in Python (not an MS technology) was written so that certain things are only allowed to be done as the cluster starts up. If it were run on linux or unix it would be exactly the same because it's not a system software level problem.
Hundreds of thousands of systems running SQL Server and Windows Server can transparently handle multiple hardware failures and the addition of new hardware without any downtime, because the software on top of them was written to allow it (typically by using an MS technology like .NET).
As for two whopping big UNIX box's, which ones exactly? There are no such computers in existence, there would be less than 100 clusters more powerful than the Eve cluster, in the WORLD. And most of them deal with physics simulations that cope just fine with crashing because you can just restart the simulation.
Someone posted this already: http://www.codinghorror.com/blog/archives/001198.html but that's why license fees mean nothing. They're significantly cheaper than the alternative, which is extensive programming time, but I get this odd feeling that like many manufacturers, in the 'mined minerals are free' crowd, many unix/linux evangelists regard time as a free commodity, because that's the only way 'free' software will win in any task requiring extensive development work.
|
|
|
|
|
| |
|
| Pages: 1 2 [3] 4 5 6 :: one page |
| First page | Previous page | Next page | Last page |