Tim Cutts wrote: > > On 19 Jan 2006, at 6:48 pm, George Magklaras wrote: > >> >> >> Tim Cutts wrote: >> > > > The processor isn't what gives SGI a niche, as you say; you can buy > less outrageously expensive Itanics from elsewhere, if you wish. What > SGI have that is, at present, more or less unique, is the ability to > build machines where you can scale CPUs, memory and I/O independently, > and which you can scale essentially as far as you like until your money > runs out. This can be very important for some niche applications. > > You want a machine with only 4 CPUs but a terabyte of RAM? You can buy > an Altix in that configuration. I don't know of anyone else evenly > vaguely mainstream that offers machines with as much physical memory in > a single system image, although please enlighten me if there is one. > We have an Altix here with 4 CPUs and 192GB of RAM - not nearly as big > as it could be, but still a configuration impossible to buy from anyone > else, at the time (and still is, as far as I know). I'd rather it > wasn't IA64, but it's the memory that's important to us, not the CPU > technology. And it's the memory that makes them expensive anyway. 48 > x 4GB DIMMs are going to be expensive regardless of whether you buy > them from SGI or anyone else. That RAM alone is around $50,000 at > current prices (http:// > osnews.pricegrabber.com/search_getprod.php/masterid=634351) I too do not know of a system config that would give you 4 CPUs and 1 Tbyte of RAM and I do not understand why one would like to stretch the CPU, memory and I/O dependency to certain limits (if you have a huge data set that you would like to place in RAM, why is it optimum to place it in 2 Tbytes of RAM and have only 4 CPUs to crunch and not 8...?), especially if you are a business and you make money. Well balanced systems do not dictate 1 CPU accessing 4 Tbytes of RAM. Although I do know of machines that could give you 4 CPUs and around half a Tbyte of RAM a couple of years ago, which is, well scalable by many folks standards (based on the p5-570 memory controller that IBM's POWER 5 570 boxes wear) and does fit the current profile of your Altix. But my point is that SGI boasts its kernel scalability features by quoting the X and the Y tests and not the independent scaling as you present it above. Their SMP scalability is of course a result of tweaks they have performed, based AND on NUMA/SMP scalability techniques pioneered by many other earlier OS platforms and Linux itself. As for the memory scalability. 64 bit processor addressing on the makes the theoretical memory addressing limit to 16 exabytes. Which means that what limits the addressable RAM is the chipset limits, which comes down to the memory controller design (bus arbitration, number of slots, etc). A major manufacturer has the technical means to scale that on the "motherboard" backplane, provided that they are going to make money selling them. You can then say that they offer indeed a unique config, but that is a result of eyeing a market, not that of technological innovation. The ability to offer a unique config based on existing technologies (what you are saying) differs from the ability to offer a unique technology(what I am talking about). > > The NSA and their friends do have the big budgets you mention, and they > want to search large databases as fast as possible. That means holding > the whole database in memory. You also need to load that database into > memory initially very quickly; the Altix can win there too - you just > keep adding HBAs and spindles until the aggregate bandwidth reaches > your requirements. We've heard of examples where SGI have built > systems to load terabyte databases into RAM from storage in a couple of > seconds. > > This is a market no-one else is really playing in, and because the > likes of the NSA and GCHQ like this sort of machine, they're not going > to let the one company go under, that can provide them. I can tell you of a branded manufacturer who did provide large RDBMS NUMA boxes, and does (AFYIK) provide the same service to people like the folks you mention off list (I do not mean to promote companies). Now, apart from RDBMS, the organizations you mention do work that requires trivial parallelization and hence they need smaller memory chunks but massive throughput. I am not sure that huge single image RAM chunks and a small number of CPUs fit that sort of workload :-). > > PS. I realise this is bioclusters, not comp.sys.sgi.advocacy, so I > apologise for the off topic post - if anyone wants to continue this > particular discussion we should take it off list. > Quite right. I think it should be clear by now that I am not an SGI advocate :-). -- -- George B. Magklaras Senior Computer Systems Engineer/UNIX Systems Administrator The Biotechnology Centre of Oslo, University of Oslo http://www.biotek.uio.no/ EMBnet Norway: http://www.biotek.uio.no/EMBNET/