Date: Sun, 04 Jan 1998 23:11:15 -0400 From: =?iso-8859-1?Q?=22Luis_E=2E_Mu=F1oz=22?= <lem@cantv.net> To: dyson@FreeBSD.ORG Cc: lukas@design.de (Lukas Wunner), andreas@klemm.gtn.com, lukas@design.de, lem@cantv.net, freebsd-isp@FreeBSD.ORG Subject: Re: [fbsd-isp] Designing for a very large ISP Message-ID: <3.0.5.32.19980104231115.0085f740@pop.cantv.net> In-Reply-To: <199801042039.PAA01086@dyson.iquest.net> References: <19980104174553.57475@reactor>
next in thread | previous in thread | raw e-mail | index | archive | help
At 03:39 PM 04/01/1998 -0500, John S. Dyson wrote: >Lukas Wunner said: >> performance, especially the VX chipset). Now consider that we'd put >> say, 512MB of RAM in the PPro based system, so that's twice as much as >> in our Pentium based news box. If there is more real memory, there is >> of course also more data to be transferred between the CPU and the memory >> per second, but as the PPro boards' memory bandwidth is actually smaller >> than the Pentium boards', the memory interface will become a real >> bottleneck. As I said, OrionGX boards seem to be a lot better in that >> respect, but they are also hard to get. At least that is my experience. >> >I have done some analysis on PPro memory usage, and one thing that is >fairly impressive is that the processor is very well decoupled from memory >by the L1/L2 cache. I think that you'd find that a PPro bus is quite a >bit less occupied than a typical P5 MB (due to a dual bus, cache/memory) >approach. A normal dual P5 will pretty much fully occupy it's memory bus, >while a dual P6 will often not. The decoupling appears to be applicable to >the PII also, due to it's dual bus architecture, even though the L2 cache is >slower than on a P6. > >That isn't to disagree with you about the lower Natoma memory bandwidth, >however, the PPro doesn't need quite as much of it, while leaving more >for peripherals. Note that writes happen a lot less often on a PPro/PII, >due to the writeback caching, while the write performance of Natoma is >where it is most deficient. This (very informative) discussion reminds me one of my favorite arguments about using PCs, which I would like you all to comment. I think that the servers at an ISP are needed only to copy data (using certain protocols) among the network and the disks. This is essentially what they do, so I'm not so concerned about the CPU power that sales-people tend to emphasize. If we get a high end server, for instance, an Alpha Server, we get a PCI bus (same bus on high end PC servers, probably at the same speed?). Sun is also in the works for a PCI Ultra (or so have I heard). Some high end workstations use 70ns memory, the same (or even slower) than the memory we place at the PC servers. The disk subsystems are built of the same hardware and has the same specs. I know there are significant differences in the internal architecture of those servers and the PCs we use, however I tend to believe they provide no significant advantage in our bussines. To ilustrate this, I currently run two POP3 servers. One runs on an high-end alpha server, the other on a Pentium 100. Both of them provide exactly the same performance. All of this adds up to my belief that for customers behind 56K or even 128K modems there are no diferences if the services are delivered by crays or by cheaper PCs. Customers behind fatter pipes ussualy have their own servers anyway ;) I will be glad to see comments on these ideas... -lem
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?3.0.5.32.19980104231115.0085f740>