| Author |
Topic |
Tar om
Minmatar
Octavian Vanguard
|
Posted - 2008.02.22 15:17:00 - [ 181]
So, any more news from CCP on this. Cos it would be a nice read during gate jumps :) |
Sazumaan Johnza
Minmatar
Capital Construction Research
Pioneer Alliance |
Posted - 2008.03.07 00:14:00 - [ 182]
Bump, in response to Attonasi's post here and because this should be the #1 priority at CCP. Any news CCP, we are all ready for Infiband...bring it on soon(ish) as Winterblink would say! |
Freaky Eyes
|
Posted - 2008.03.07 14:48:00 - [ 183]
Does anyone know what hardware the current sol nodes are running on?
AMD dualcores i thought, but what kind? |
Sherrisa
ANZAC ALLIANCE
Southern Cross Alliance |
Posted - 2008.04.02 02:31:00 - [ 184]
Bump, any further news on this? |
StoreSlem
Minmatar
4S Corporation
Morsus Mihi |
Posted - 2008.04.02 03:41:00 - [ 185]
My idea: slow down the nodes that are under heavy load. Make whatever's happening, happen at a manageable rate for the node. Slows the game but improves response times, a game in this state would very much be playable. Then, fix the desyncs. |
Ruah Piskonit
Amarr
PIE Inc.
|
Posted - 2008.04.02 06:56:00 - [ 186]
|
Mutabae |
Posted - 2008.04.02 07:03:00 - [ 187]
 Originally by: StoreSlem
My idea: slow down the nodes that are under heavy load. Make whatever's happening, happen at a manageable rate for the node. Slows the game but improves response times, a game in this state would very much be playable. Then, fix the desyncs.
ehh? the node is already 'managing' the load... the load is already stacking up player commands until it can process them... the node is already 'fine' (as fine as it can be when its cpu is maxed out 100% and 5-15 minutes behind what you are telling your client to do...). Slowing the client down would do ... what? "You command is in a queue... estimated time to execution is now... 6.32 minutes" ? Seriously, where's the 'playability' enhancement going to come from? The node's *already* maxed out. Until >1 cpu is applied to a node, there can't be a benefit to you as far as responsiveness goes. |
Semkhet
Dark Tornado
Ethereal Dawn |
Posted - 2008.04.02 07:51:00 - [ 188]
Edited by: Semkhet on 02/04/2008 07:55:07
I've no crystal ball, but I've the feeling that CCP is struggling with some fundamental assumptions & decisions which have hampered their ability to implement a scalable and fault-tolerant system.
In my case, I often use OpenVZ virtualization as soon my client requires multiple servers needing isolation. Apart being free and giving you access to source code, OpenVZ supports 0-downtime migration of RUNNING machines between hardware nodes. Another benefit is that since OpenVZ binds intimately with the underlaying OS, the virtualization in itself uses no more than 3% of the hardware node resources.
I frankly never understood why CCP went the Microsoft way, at a time when governments representing more than 3 billion people on earth have come to the conclusion that Linux is best the way to safeguard their interests in the long run in terms of control, stable in-house knowledge capitalization, independence, performance and costs.
Since 2005, India alone puts on the market 300'000 IT graduates each year, a number similar to the whole population of Iceland.
Let's face it: CCP chose to do something in a way which was never attempted before in a live commercial production environment, but implemented it with OS & tools they can't modify without entering in heavy non-disclosure agreements and all that crap. In fact, software-wise, their control is probably limited to their application, while they rely heavily on the collaboration of third-party companies to solve problems appearing in the underlying layers (drivers/SQL/OS) since they have no easy way to fiddle with the source code of these layers, nor is it obvious that the enjoy the related knowhow touching such different realms.
Would they have based their product on open-source technologies, they could hire an army of IT engineers in emergent markets and get research/development done for peanuts without facing the nightmare of finding legal ways to outsource development tasks on software they don't even own the rights (OS/SQL/drivers). And when you really need mathematicians, you turn to Russia where math PhD's holding academic chairs cost €600.-/month...
If you take patent statistics to define if a given country enjoys a good development environment (remember that a patent according to WIPO definitions should exhibit novel properties), we find that Iceland holds the impressive record of being the second highest patent filer in Europe for... biotechnologies. But in IT, they fare not better than Slovenia, a country which is not precisely an IT engineer's dream from my standards last time I visited the place.
Add to that CCP's requirement that their devs must relocate to Iceland, and how many collaboration venues are you preempting ? That in the epoch in human history where decentralized/distributed development, research and production has never been easier ?
I sometimes bang my head trying to make sense of all this. It just doesn't add up. |
Mr Friendly
The Lost and The Damned |
Posted - 2008.04.02 08:02:00 - [ 189]
Edited by: Mr Friendly on 02/04/2008 08:13:28 Originally by: Semkhet
I've no crystal ball, but I've the feeling that CCP is struggling with some fundamental assumptions & decisions which have hampered their ability to implement a scalable and fault-tolerant system.
In my case, I often use OpenVZ virtualization as soon my client requires multiple servers needing isolation. Apart being free and giving you access to source code, OpenVZ supports 0-downtime migration of RUNNING machines between hardware nodes. Another benefit is that since OpenVZ binds intimately with the underlaying OS, the virtualization in itself uses no more than 3% of the hardware node resources.
I frankly never understood why CCP went the Microsoft way, at a time when governments representing more than 3 billion people on earth have come to the conclusion that Linux is best the way to safeguard their interests in the long run, both in terms of independence, performance, capitalization and costs.
Since 2005, India alone puts on the market 300'000 IT graduates each year, a number similar to the whole population of Iceland.
Let's face it: CCP chose to do something in a way which was never attempted before in a live commercial production environment, but implemented it with OS & tools they can't modify without entering in heavy non-disclosure agreements and all that crap. In fact, software-wise, their control is probably limited to their application, while they rely heavily on the collaboration of third-party companies to solve problems appearing in the underlying layers (drivers/SQL/OS) since they have no easy way to fiddle with the source code of these layers, nor is it obvious that the enjoy the related knowhow touching such different realms.
Would they have based their product on open-source technologies, they could hire an army of IT engineers in emergent markets and get research/development done for peanuts without facing the nightmare of finding legal ways to outsource development tasks on software they don't even own the rights (OS/SQL/drivers). And when you really need mathematicdians, you turn to Russia where math PhD's holding academic chairs cost €600.-/month...
If you take patent statistics to define if a given country enjoys a good development environment (remember that a patent according to WIPO definitions should exhibit novel properties), we find that Iceland holds the impressive record of being the second highest patent filer in Europe for... biotechnologies. But in IT, they fare not better than Slovenia, a country which is not precisely an IT engineer's dream from my standards last time I visited the place.
Add to that CCP's requirement that their devs must relocate to Iceland, and how many collaboration venues are you preempting ? That in the epoch in human history where decentralized/distributed development, research and production has never been easier ?
I sometimes bang my head trying to make sense in all this. It just doesn't add up.
From the Dev posts I have read, the scalability issue comes down not to their use of MS server software for DBs etc, but CCP's reliance on open-source StacklessPython for their node-load management. Their server software simply can't apply more than one cpu per node, iirc. That wasn't an anticipated need when they built the code, I guess. Poor planning, but hindsight bias is 20/20. As I'm not in any way a coder, I can't knowledgeably criticize their choice, though I do have the feeling that if they temp-hired some Indian or Asian programmers for a few months, the problem would just go away. Those countries seem to ooze underutilized, inexpensive and talented programmers, while CCP seems to be stuck in rut. 500 Korean MMOs that run great, play 'for-free' and have great graphics while running on low-quality hardware can't all be wrong, can they?  I hope CCP does fix the cpu<->node issue, or Eve's support simply can't improve to match its vision... God-sized pipelines between cpus won't help if the processors themselves are the chokepoints... |
Raem Civrie
Deep Core Mining Inc.
|
Posted - 2008.04.02 08:46:00 - [ 190]
Originally by: Mashie Saldana
The main point of Infiniband isn't for RAMSAN connectivity but for node to node connectivity, I have a feeling that one already is connected to the database servers using fibre channel.
The main reason to use Infiniband isn't the bandwith but the very low latency compared to Ethernet. Due to that RDMA (Remote Direct Memory Access) is possible, as in node A can read directly from node B's RAM about the status of say a ship. This is essential to be able to break the one node per Sol limitation. Now however just ripping out all Ethernet cards and plugging in Infiniband cards instead won't do much good until more or less all code is rewritten to use the new abilities.
So as Soon™ as the coding is done I'm sure CCP will deploy it.
I'm pretty sure none of that was in english. Same goes for the rest of this thread. I knew I should've finished that CCNA. |
Semkhet
Dark Tornado
Ethereal Dawn |
Posted - 2008.04.02 08:54:00 - [ 191]
Edited by: Semkhet on 02/04/2008 08:59:49 Originally by: Mr Friendly
From the Dev posts I have read, the scalability issue comes down not to their use of MS server software for DBs etc, but CCP's reliance on StacklessPython for their node-load management. Their server software simply can't apply more than one cpu per node, iirc. That wasn't an anticipated need when they built the code, I guess.
As I'm not in any way a coder, I can't knowledgeably criticize their choice, I still do retain the feeling that if they temp-hired some Indian or Asian programmers for a few months, the problem would just go away. Those countries seem to ooze underutilized, inexpensive and talented programmers, while CCP seems to be stuck in rut.
I hope CCP does fix the cpu<->node issue, or Eve's support simply can't improve to match its vision... God-sized pipelines between cpus won't help if the processors themselves are the chokepoints....
Yes, I'm aware of their position, but from my perspective, where they have made a structural nonsense decision is that you don't try to do something new choosing a model where your " absolute control & knowledge" is limited to the application layer. I can give an example: A few years back I get a contract where a shipyard wants to implement a distributed worker's real time task control in their hangars to know who's doing what, where and when. The problem is that these hangars use heavy machinery which creates an extremely polluted EM environment. Hence I choose the microcontroller the slaves will be based on, design the mainboards and create an asynchronous master-slave networked system, where a server pool slaves to retrieve information every few minutes, processes the data and serves a synoptic & detailed view of everything which is going on to the direction. The least of the problems here is the final application. Given the EM interferences showering the network, I program a transmission protocol based on non-blocking calls, including extensive error-checking and compression. The microcontrollers are shipped with their official network library providing all the core functions my protocol will use. Cool I think, no no need to reinvent the wheel at such a trivial stage. So everything gets implemented, the first 10 prototypes built, installed, and the testing begins. And what happens ? Randomly, the RTC clocks of the slaves deliver a completely incoherent time data. The funny thing is that the RTC clocks still hold the right time even when spitting out nonsense. All code and layers are reviewed separately and not a single bug is detected, simulations show that everything should work fine. Yet it isn't the case. After 3 weeks of frustration, I finally decide to take apart the only bit we didn't write in-house: the network library of primitives. Disassemble it, recompile the assembler code with debugging symbols, hook the whole sheebang to a logic analyzer and map the activity to the schematics of the microcontroller. And what do we find ? That the geniuses which programmed the primitives never bothered to consider that the onboard cloak could be used simultaneously to the network: one of the I/O pins used by the network driver is shared with the I2C BUS, which in turn is connected to the RTC. So what was happening is that since tasks where occurring concurrently and asynchronously in the slaves, any slave interacting with the network exactly at the same time a dialog was going on between its CPU/RAM and the RTC would interfere with the flux of bits coming out of the RTC eeprom chip turning it into pure garbage. That's why it is important to master every single aspect of what you're doing if you don't want to depend of third-party companies (and the associated costs...). |
StoreSlem
Minmatar
4S Corporation
Morsus Mihi |
Posted - 2008.04.02 09:48:00 - [ 192]
Originally by: Mutabae
Originally by: StoreSlem
My idea: slow down the nodes that are under heavy load. Make whatever's happening, happen at a manageable rate for the node. Slows the game but improves response times, a game in this state would very much be playable. Then, fix the desyncs.
ehh? the node is already 'managing' the load... the load is already stacking up player commands until it can process them... the node is already 'fine' (as fine as it can be when its cpu is maxed out 100% and 5-15 minutes behind what you are telling your client to do...). Slowing the client down would do ... what? "You command is in a queue... estimated time to execution is now... 6.32 minutes" ?
Seriously, where's the 'playability' enhancement going to come from? The node's *already* maxed out.
Until >1 cpu is applied to a node, there can't be a benefit to you as far as responsiveness goes.
No, as in run the game at half speed, it would be able to manage twice the numbers of players. |
Dave3of5 |
Posted - 2008.04.02 12:11:00 - [ 193]
Originally by: Mr Friendly
Edited by: Mr Friendly on 02/04/2008 08:13:28
From the Dev posts I have read, the scalability issue comes down not to their use of MS server software for DBs etc, but CCP's reliance on open-source StacklessPython for their node-load management. Their server software simply can't apply more than one cpu per node, iirc. That wasn't an anticipated need when they built the code, I guess. Poor planning, but hindsight bias is 20/20.
As I'm not in any way a coder, I can't knowledgeably criticize their choice, though I do have the feeling that if they temp-hired some Indian or Asian programmers for a few months, the problem would just go away. Those countries seem to ooze underutilized, inexpensive and talented programmers, while CCP seems to be stuck in rut. 500 Korean MMOs that run great, play 'for-free' and have great graphics while running on low-quality hardware can't all be wrong, can they?
I hope CCP does fix the cpu<->node issue, or Eve's support simply can't improve to match its vision... God-sized pipelines between cpus won't help if the processors themselves are the chokepoints...
No offence to you or Semkhet but when anyone says to me about outsourcing programming to Indian, Russia ... etc I cry. Outsourcing the programming of this game to some cheap development house in India would destroy this game (I know I've seen the results of outsourcing programming). Those Korean MMOs are a totally different ball-game and allot of them still have tons of bugs / design problems … etc Also throwing programmers and Mathematicians at a problem would not solve the problem I know your not a programmer but any "programmer" that suggests adding some huge development team to an already existing project is very miss-informed. First year in my University course covered this as a Computer Science / Software Engineering basic. I thought at the time it was bull but in the real-world you can see why. What they are doing is fine any more comments on prgramming have to be ignored really. |
Semkhet
Dark Tornado
Ethereal Dawn |
Posted - 2008.04.02 12:21:00 - [ 194]
Originally by: StoreSlem
No, as in run the game at half speed, it would be able to manage twice the numbers of players.
You are aware that such a solution would be pure duct-tape style while addressing none of the core problems right ? IMHO, I can't imagine a project like this MMO not being based on an application-independent clustered virtual machines pool enjoying automatic load-balancing properties and on-the-fly relocations. The technology exists, works, and is free. Then we can start talking about everything else, and they can program their game with whatever language suits them. Besides, when talking about Stackless Python, the main point is not that's Open Source, but that it's considered EXPERIMENTAL even by their makers. In fact if you visit the Stackless web site, you won't find mention of more than 4 commercial applications (EvE included), and 4 non-commercial apps. So not a lot of people are coding with SP, hence the established community's knowledge base you can cater on when facing bugs & problems is extremely reduced compared to more classic languages. Not a cautious choice when going live with a commercial product under constant development. Would I have to lead EvE's implementation, I would certainly not choose as main language one of the least used ones just under pretext that it might shorten the design cycle of game features, while happily forgetting the time you will spend debugging when things don't work as expected taking into account that there's almost nobody who can help you. Just check Stackless mailing list archives: a pathetic 50 messages for the whole month of March 2008... |
Semkhet
Dark Tornado
Ethereal Dawn |
Posted - 2008.04.02 12:40:00 - [ 195]
Edited by: Semkhet on 02/04/2008 12:42:01 Originally by: Dave3of5
No offence to you or Semkhet but when anyone says to me about outsourcing programming to Indian, Russia ... etc I cry.
Outsourcing the programming of this game to some cheap development house in India would destroy this game (I know I've seen the results of outsourcing programming). Those Korean MMOs are a totally different ball-game and allot of them still have tons of bugs / design problems … etc
Also throwing programmers and Mathematicians at a problem would not solve the problem I know your not a programmer but any "programmer" that suggests adding some huge development team to an already existing project is very miss-informed. First year in my University course covered this as a Computer Science / Software Engineering basic. I thought at the time it was bull but in the real-world you can see why.
What they are doing is fine any more comments on prgramming have to be ignored really.
I can tell that outsourcing is an art which only comes into play when you have a pretty precise idea about what you are doing. You can outsource almost everything as long you do it by isolated bits, because you're the only one who wants to retain the knowledge concerning how everything fits together. And who ever told that by outsourcing you loose control about the quality and structure of your product ? I'm in the IT field since early 80's, and I can tell not only that outsourcing works, but that on top in this new globalized world, you better master outsourcing if you want to maximize your chances to stay in business on the long run. The first ones that started massive IT outsourcing have been swiss banks and insurances, they do that since mid-80's and guess what ? They're nowadays among the first 10 biggest & richest players of their field worldwide. The only entities which have no need whatsoever to outsource abroad are government-related agencies which excel in wasting taxpayer money by using it as hidden subventions for local industries. Besides, russian intellectuals are among the most productive people I ever had the privilege to work with as soon they get respect, a suitable work environment and correct funding. |
N0R0X
S.K. Astronautics
Cascade Associates |
Posted - 2008.04.02 12:57:00 - [ 196]
Hope my addition is seen as constructive. I have seen many threads that I think I have not replied to (thank goodness since they eventually turned into flameboats) where peeps keep saying "Why is CCP using crappy MS-SQL when MySQL is faster". Well this is actually a very complex explanation, so I will give a very short version for the non-technical to understand. MySQL in the beginning was just not a contender for the MS-SQL. It lacked features, speed and reliability that MS-SQL had, and remember that CCP started beta EVE around 2001-2003. Now if you were a company that needed a robust database system and you had the choice between MS (Running for ages) and MySQL (Still gathering pace) which would you pick? I would go with the reliable robust system even if I had to pay for it. It’s business sense. Ok well now you say “Well why is CCP not moving over to MySQL then since its better way now” Two things: 1. It’s not that easy to just wave a wand and all now works with the new system. CCP are already going through massive code changes to implement their future InfiniBand system. Moving all data to a different database isn’t just that. Lots of game code will need to be changed. 2. Although MS-SQL is still a little slower than MySQL, MySQL still lacks features that the MS version has. Whether CCP use them I don’t know. Next bit is where peeps always say the nodes are not fast enough and we need to give each system its own node. One word “Jita” Yup you guessed it. Jita apparently already has its own node but remember that processors currently contribute very little to lag. Unless of course they are at 100% load and start to cause queuing. I can’t find the post but I remember CCP replying to a post a year ago about the servers are below 50% load. So that tells me and hopefully you the problem is not the processors. Ok CCP are upgrading the systems but that because they are implementing massive data pipes for which maybe the old hardware would bottleneck. For a game like EVE data needs to be moved at high speed and not necessarily high volume (hope I’m correct here) so it means you need ultra low latency hardware. Finally, I am not trying to cause a flamebait/boat just trying to explain in nice terms maybe why CCP has not moved in a certain direction. When it comes to ideas, although some already have, give constructive ones like how to combat lag specifically what could be changed. I’m sure if CCP thinks it might work that they will look into it. Don’t just throw at the forums and CCP “Just get new OS’s/Databases/Trillion $ hardware” because things don’t simply just go together like Lego. As for the HPC progress, yes it would be very nice to hear some more info so that I can awe and drool over it.  If you think I am trolling then delete my post dev's/gm's Happy EVE’ing everyone. |
AleRiperKilt |
Posted - 2008.04.02 13:09:00 - [ 197]
You CCP guys should try Oracle... seriously... give it a shot. Heck, I'll even help you tweak the queries ( for a nominal ISK fee  ) Just ask around big corporations and banks, which database they trust their million of transactions a day? Besides custom COBOL apps, it's between Oracle and DB2. |
Nareg Maxence
Gallente |
Posted - 2008.04.02 13:12:00 - [ 198]
Discussion thread participator breakdown:
10% Totally clueless people. (Semkhet, try to restrain yourself, please..)
70% Clueless people who springle buzzwords around for credibility.
10% Non-clueless people with useless ideas.
10% Who actually have some clue.. |
Semkhet
Dark Tornado
Ethereal Dawn |
Posted - 2008.04.02 13:39:00 - [ 199]
Originally by: Nareg Maxence
10% Totally clueless people. (Semkhet, try to restrain yourself, please..)
What you say dude, my wallet says otherwise, and I'm not talking about the ingame one  |
Zeba
Minmatar
Honourable East India Trading Company
|
Posted - 2008.04.02 13:51:00 - [ 200]
I already know what CCP plans to do with Infiniband and its potential for mostly killing lag even for fleet battles if its coupled with enough new servers getting added to the 0.0 regions. I simply want an involved Dev to give us a short paragraph of its current state and if its looking good so far.  |
Vladimir Ilych
Gradient
Electus Matari |
Posted - 2008.04.02 13:55:00 - [ 201]
Originally by: Semkhet
The first ones that started massive IT outsourcing have been swiss banks and insurances, they do that since mid-80's and guess what ? They're nowadays among the first 10 biggest & richest players of their field worldwide.
Not on topic but.. Swiss bank UBS writedown |
Tzrailasa
Tzrailasa Corp
|
Posted - 2008.04.02 14:03:00 - [ 202]
Originally by: Zeba
I already know what CCP plans to do with Infiniband and its potential for mostly killing lag even for fleet battles if its coupled with enough new servers getting added to the 0.0 regions.
According to what CCP has said as to how they're going to use this, it'll NOT mean any significant decrease for lag in fleet-battles in itself. Fleet-battles on the same grid will still be processed by only one CPU, and since the problem causing lag in battles appear to be the CPU having reached its limits, there'll be little to no effect on fleet battles. They MAY be able to offload a few things to other processors, like info lookups and chat, but things like module lag and load lag will most likely stay the same. New servers will not help much either..... Where this'll make a huge difference is in places like Jita and high-load mission systems. These systems currently run on only one CPU and when that reaches its limit everyone in system lags. After InfiniBand CCP will be able to assign several CPU's to each system (maybe even dynamically), meaning lag in these systems will most likely either disappear or at least become much less. All in all, this is a good addition for EVE as such, but it'll (as far as I understand it) not help us long-suffering fleet combatants much...  |
Zeba
Minmatar
Honourable East India Trading Company
|
Posted - 2008.04.02 14:06:00 - [ 203]
Edited by: Zeba on 02/04/2008 14:09:04 Originally by: Vladimir Ilych
Originally by: Semkhet
The first ones that started massive IT outsourcing have been swiss banks and insurances, they do that since mid-80's and guess what ? They're nowadays among the first 10 biggest & richest players of their field worldwide.
Not on topic but..
Swiss bank UBS writedown
Yes but those server farms only have to do one function. Process banking/insurance transactions and keep track of them which anyone with half a brain can be trained to work on due to the completely unchanging nature of how finance functions between the institutions that use them. Do you really think they are coded to function as an unsharded gaming world with the need to resolve hundreds of thousands or even millions of calculations per second just off one fleet battle and still keep up with all the other processes going in the game? Sure banking and rl corporate server clusers are reliable just as the eve servers would be reliable if it only had the market to deal with and not the gaggle of other areas that need its own custom code for. Eve gaming universe server cluster =! rl financial server cluster. |
Zeba
Minmatar
Honourable East India Trading Company
|
Posted - 2008.04.02 14:14:00 - [ 204]
|
Burnharder |
Posted - 2008.04.02 14:38:00 - [ 205]
Originally by: Zeba
I already know what CCP plans to do with Infiniband and its potential for mostly killing lag even for fleet battles if its coupled with enough new servers getting added to the 0.0 regions. I simply want an involved Dev to give us a short paragraph of its current state and if its looking good so far.
Here it is: void ServerNode::DoInfinibandThing(shared_ptr<LotsOfEntities> theEntities) { // TODO - Infiniband thing } |
Zeba
Minmatar
Honourable East India Trading Company
|
Posted - 2008.04.02 14:52:00 - [ 206]
hehehe. ^^ |
StoreSlem
Minmatar
4S Corporation
Morsus Mihi |
Posted - 2008.04.02 15:25:00 - [ 207]
Originally by: Semkhet
Originally by: StoreSlem
No, as in run the game at half speed, it would be able to manage twice the numbers of players.
You are aware that such a solution would be pure duct-tape style while addressing none of the core problems right ?
Of course, it's just a dynamic security net to smooth the load peaks, which comes in addition to any other improvements. Any organization with aspirations of profit knows it's even worse running below full capacity than above it, which is why this would be a viable alternative. Put it bluntly, there is no way CCP will give us a server park that can handle the peaks as that would mean 99% of the time they have large amounts of idle resources. |
Seth Ruin
Minmatar
Ominous Corp
Circle-Of-Two |
Posted - 2008.04.02 15:39:00 - [ 208]
Originally by: StoreSlem
Put it bluntly, there is no way CCP will give us a server park that can handle the peaks as that would mean 99% of the time they have large amounts of idle resources.
They could run SETI@Home in the background? |
Zeba
Minmatar
Honourable East India Trading Company
|
Posted - 2008.04.02 15:41:00 - [ 209]
Originally by: StoreSlem
Originally by: Semkhet
Originally by: StoreSlem
No, as in run the game at half speed, it would be able to manage twice the numbers of players.
You are aware that such a solution would be pure duct-tape style while addressing none of the core problems right ?
Of course, it's just a dynamic security net to smooth the load peaks, which comes in addition to any other improvements. Any organization with aspirations of profit knows it's even worse running below full capacity than above it, which is why this would be a viable alternative. Put it bluntly, there is no way CCP will give us a server park that can handle the peaks as that would mean 99% of the time they have large amounts of idle resources.
Thats where a subscription drive comes in. Currently CCP has little to no advertising so as to keep the new subs joining at a sustainable rate for overall server stability. To me seeing them expand Eves direct marketing visibility onto Steam is only a good sign that the stuff needed to buff the cluster and run mostly lag free for all players truely is coming soon. Moar subs = Moar money to CCP = Moar yarrdware to make it run really well. Remeber that most players complaining about lag are fleet battle guys who share one cpu with an entire region of space. If that changes to one cpu per 2 or 3 systems instead of a whole region then even without infiniband I think you might see greatly reduced lag for the majority of fights. But thats takes money to install and money to keep running. And as long as CCP is turning a profit why the hell would they care if at times the cluster runs at a low percentage of useage? Trust, they make alot moar profit that you think they do. |
Blinky M'eizrtard |
Posted - 2008.04.02 19:10:00 - [ 210]
Originally by: Eigenvalue
Finally, for the person who suggested the nVidia computation framework, this is an interesting way to deal with it, but I'm not sure how well it'll work. Generally GPU computations can't do random numbers well, and at least the way it was when I did this in the past they didn't deal with branches well either. It's really good for long deterministic calculations like some scientific simulations, etc, but Eve is essentially a big montecarlo simulation, which AFAIK don't work well on GPU based platforms.
I dunno which computations are eating the most CPU at the moment, but perhaps a local coprocessor could take load off the main processor. FPGAs are pretty good at Monte Carlo-class problems and use little power. You can get FPGA boards that plug in to existing CPU sockets like AMD Socket F and share the onboard low-latency interconnect (e.g., HyperTransport). You can program FPGAs with high-level languages--and would need to, since some recoding is definitely involved. Like with GPGPUs, you have to pick and choose which parallel problems to shift to the FPGA coprocessor. Branching code is not so good, but you can do a fast Mersenne Twister RNG in an FPGA. N-body simulation has also been done, with good performance. But if the real bottleneck is internode I/O latency, then either InfiniBand or some rearchitecting of the cluster is probably needed. I'd sure be interested to read about how the Lag Slayer works when it's deployed! |
|
All Channels
EVE General Discussion
