Date: Tue, 5 Nov 1996 18:30:45 -0600 (CST) From: Joe Greco <jgreco@brasil.moneng.mei.com> To: richardc@CSUA.Berkeley.EDU (Veggy Vinny) Cc: jgreco@brasil.moneng.mei.com, michaelv@MindBender.serv.net, mark@quickweb.com, terry@lambert.org, imp@village.org, jmb@freefall.freebsd.org, isp@freebsd.org, chad@gaianet.net, mario1@PrimeNet.Com, johnnyu@accessus.net Subject: Re: /usr/obj size Message-ID: <199611060030.SAA07510@brasil.moneng.mei.com> In-Reply-To: <Pine.PTX.3.95.961105130741.13095P-100000@soda.CSUA.Berkeley.EDU> from "Veggy Vinny" at Nov 5, 96 01:20:24 pm
next in thread | previous in thread | raw e-mail | index | archive | help
> > Vince, > > > > IMHO it depends on what you are going to do with the machine. Are you going > > to have lots of readers? Lots of inbound and/or outbound feeds? (Remember > > that many sites these days feed via "innfeed", and this counts as multiple > > feeds)... > > All the users will be using nntp from within GAIANET.NET. We will > have a newsfeed from Concentric Network but will probably have some feeds > to others. How MANY users? If you have no idea, project 1/10 of your number of modems. That is an almost useless guess, but it is at least probably going to be the right order of magnitude. > > If you are going to have one user reading news and have a single inbound > > feed, sorted, held for half a week, and nothing else, you might be able to > > do it on a fairly skimpy system. > > > > My idea of fairly skimpy would be 486DX/133, 64MB RAM (maybe 48MB), a fast > > disk for root/var/newslib, and a big disk for spool. You might be able to > > get away with less memory by using C-News. > > We'll have this as the minimum of a sysem. Dual PentiumPro-200Mhz > with 128MB of ram and a Seagate Elite 9 9.1 gig Fast Wide SCSI-2 drive to > start with. Dump the Dual, dump the Pentium Pro. You are not doing raytracing, you are providing news service. For this kind of application you need to be able to move data quickly. Get a P133 on a good Triton-II board. Use the money you save to buy more disks. With PP200's going for over $1000 each, you should be able to get half a dozen 1G disks with the savings. (You do NOT want to use the Elite on a machine with multiple news feeds. You do NOT want to use the Elite on an ISP class news server. It will melt. No way in hell can it handle the number of transactions per second required!) Get more RAM. Then get more CPU. In that order! :-) > > My standard cuff calculation formula for RAM is: > > > > 1MB * active readers + 2MB * number of feeds + 8MB for system + 2 * > > sizeof(history.pag) > > > > I consider this to be a minimum. News will make very good use of as > > much memory as you can give it. One of my clients considered it to be > > an excessive maximum and learned a very expensive lesson when their > > news service sucked. > > Hmmm, when you mean by active readers, you are talking about the > amount of nntp clients accessing the newsserver at once? Yes. "active" refers to any client that has not been sleeping for 15 minutes... you get some of those too and it is OK to let them swap out. > > For a CPU, I have not found a significant need to go beyond a P133. > > 'newspump.sol.net' is a P133, and is a dedicated feeds machine. It > > ranks #25 in the Freenix ratings this month. A large client has 100-150 > > nnrp sessions reading news on their P133 boxes, with CPU to spare. This > > game isn't really about CPU, it is more about memory, caching, and I/O. > > Besides, CPU is cheap. > > Hmmm, okay.... Do you or anyone know how well a FreeBSD machine > would compare to a SUN? I was saying that the FreeBSD machine can easily > do better than the SUN's since I can see how well wcarchive.cdrom.com is > doing with the high loads. You get more bang for your buck with the PC. CPU is 1/5 the price, memory is 1/2 the price, I/O controllers are 1/10 the price. Take the money you save and buy disks. I run news servers under Solaris. Spool.mu.edu is a doggy SPARCstation 10/30 that is roughly equivalent to a P60. It is a very busy machine.. it is a good machine... but intuitively I believe that a PC equivalent would be faster, and absolutely cheaper. > > Chipset is ultimately important. You WILL not be successful if you buy > > a cruddy motherboard/chipset. Buy Triton-II. I recommend the ASUS > > P/I-P55T2P4, or P/E-P55T2P4D (see http://www.asus.com.tw). I recommend > > buying hardware from Rod Grimes, I have never been disappointed by his > > services and support. <sales@gndrsh.aac.dev.com>. > > Yep, I know. I got a P55TP4XE from Rod and it's still working in > one piece today with no problems whatsoever. Too bad mines is the > original Triton though. Rod surely knows the part he sells and can give > you every bit of info you want to know. I am using several as well. Work great, no parity, only 128MB RAM supported, but really one of the best boards available at the time. > > Do not compromise on your I/O system. News is extremely taxing on a > > machine's disks. Buy multiple SCSI busses. It is much better to buy > > three $70 ASUS SC-200 NCR-810 controllers than one $220 AHA-2940 SCSI > > controller (but if you want to spend $660 on three AHA-2940's, that > > is not a bad solution either). Buy lots of disks. Stripe them with CCD. > > The reader machines noted above have 12 disks each: > > What about for Wide SCSI or Ultra/UltraWide SCSI drives? Which > controllers would be good? To hell with Wide SCSI. If you are transferring 50 MB of data from a large file, you will be well served by a pair of CCD'd drives with Wide SCSI and a small interleave. News transfers typically are 4K and are rarely larger than 64K. Your Wide SCSI bus will transfer 4K of data twice as fast, but "twice as fast" is really irrelevant since your process is already blocked waiting for the data, and whether it waits 20 milliseconds of seek time plus some real teeny number 'N' for the Wide SCSI transfer, or 20 milliseconds plus 2 * N, is really pretty irrelevant because N is so small compared to the 20 milliseconds you just waited for your data. You are better served by taking the money saved and buying more drives. Then you are more likely to have a lower number than "20 milliseconds" because latency is lower due to contention being lower. (The 20 number I pulled out of the air, the argument is valid but the number may not be). Exception: Wide SCSI may be useful for: alt.binaries (large articles) newslib (active file, history writes) but I do not think it is worth the effort myself. None of my news servers have Wide SCSI and one of them is in the Top 25. > > ---------------- ---------------- ---------------- > > | sd0 root | | sd10 var | | sd20 newslib | > > | 2G ST32550N | | 2G ST32550N | | CCD 1G 31055N| > > |--------------| |--------------| |--------------| > > | sd1 newslib | | sd11 nov | | sd21 nov | > > | CCD 1G 31055N| | CCD 1G 31055N| | CCD 1G 31055N| > > |--------------| |--------------| |--------------| > > | sd2 news | | sd12 news | | sd22 alt | > > | CCD 2G 32550N| | CCD 2G 32550N| | CCD 2G 32550N| > > |--------------| |--------------| |--------------| > > | sd3 alt | | sd13 binaries| | sd23 binaries| > > | CCD 2G 32550N| | CCD 4G 15150N| | CCD 4G 15150N| > > ---------------- ---------------- ---------------- > > I can see that the 2 GIG are all Barracuda's from Seagate but what > kind of drives are the others? 31055N is Hawk-2 1GB "ST-31055N" "Ultra-SCSI", goes for about $330 I believe... There is a poor cousin "31051N" version which is basically not Ultra-SCSI but a nearly identical drive. 15150N is Barracuda 4GB. There is now a low profile variant out but I have not seen them. > > Notice I stripe across controllers... also note the rants I have posted > > in the past on CCD stripe sizes. Use large ones except for the newslib > > disk. > > Hmmm, how do I do the CCD stripe? I heard you can make multiple > drives into one partition under FreeBSD. man ccd :-) I will provide examples if you get lost but right now I am busy helping one of my clients customers with a security problem... so I will say "RTFM" and search the mailing list. > > With 150 readers, this is DAMN BUSY... > > > > Rule #1) People always try to cheap out on news servers. > > > > Rule #2) They fail. > > > > Remember those rules and you have a chance of designing a good news > > service... > > Yep, that's true. Since it's better to spend more to get quality > components than getting it cheap since by the time, you find out, you need > to replace all the inferior components. Precisely... 20% more up front saves you 120% down the road. That does not make it any easier to spend the 20%, untul you start talking with someone who spent the 120% and spent several dozen engineering hours (wasted) to find out that their hardware is crud. That is expensive in other ways. ... JG
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199611060030.SAA07510>