Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!cs.utexas.edu!rice!sun-spots-request From: lmb@vicom.com (Larry Blair) Newsgroups: comp.sys.sun Subject: Re: Server Questions Keywords: Networks Message-ID: <3488@brazos.Rice.edu> Date: 2 Dec 89 17:04:10 GMT Sender: root@rice.edu Organization: Sun-Spots Lines: 72 Approved: Sun-Spots@rice.edu X-Refs: Original: v8n205 X-Sun-Spots-Digest: Volume 8, Issue 213, message 3 of 12 In article <3255@brazos.Rice.edu> you write: =X-Sun-Spots-Digest: Volume 8, Issue 205, message 13 of 19 = =We are considering a server for both NFS and compute engine use and have =some questions about resource utilization: I'll try to give you a data point. I'm running two servers right now. One is a 3/160 running 3.5.1. It has six SMD disks (4 CDC FSD's, a CDC 9720-850, and a Fuji 2361A) running on 2 Interphase 4200 controllers, one of which is using a 4-drive adapter. It also has an ALM-1 and an ALM-2 and has Telebit Trailblazer Pluses on the CPU's tty ports. It has 16 Mb of memory. The server is used as a netwide (20 workstations) fileserver, mail server, print server, and netnews server. It also has as many as 20 users logged in thru the ALM ports. It processes about 40 Mb a day of Usenet traffic thru the TB's. Due to NFS's performance problems for writes, it is also used as a compute server for all 3.5 compilations. Performance is excellent. Running a perfmeter shows that CPU usage rarely reaches 100%. The limiting factor is the I/O bandwidth to the disks, even with the fast caching controllers. Interestingly, I recently came across a paper, written by one the guys at Legato, which supported my observations. My other server is a 3/160 running 4.0.3. It has 4 Hitachi DK515-78 drives running with an Interphase 4201 controller. This turns out to be a very cost effective solution. The drives are 20mhz, 780 Mb, ESDI's and only cost $2,300 each. My biggest problem with this system is that it only has 8 Mb of memory, which is completely insufficient for a fileserver under 4.0, particularly since it is also used as a compute server for all 4.0 compilations. I have a 40 Mb 3/470 on order, which will replace it. Eventually I'll have eight disks on two controllers and, possibly, a Legato board. =1. How much does NFS activity slow compute-intensive tasks? Some, but the real effect is on overall I/O throughput. =2. What is the reasonable limit on the number of NFS mounts to a server = if there is one? We haven't hit it yet. I don't really think that there is one. =3. How does the ammount of system ram affect the NFS response time? The = ammount of swap? I'm not sure how much it actually affects NFS performance, but on systems like ours that also perform many other functions it makes a tremendous difference, particularly with 4.0. =4. Do NFS "accelerators" like PrestoServe really work? How well? Absolutely. I've observed one in action. It was in a 3/280 that was serving 200 (!!!) workstations. Despite the small cache, only 1 Mb, the hit ratio was over 90%. It transforms a server from a read-only device into a truly usable file server. =5. Is it better to have 2 small (ss330) servers, one for NFS use, and one = for CPU tasks, or one biggun (ss370 - 390) doing both? Note that a 330 and the 370,390 use the same CPU. The cost effective way is to buy the cheapest configuration that Sun will sell and then add controllers from Interphase or Ciprico. Since the problem is NFS write speed, not CPU usage, one server with a Legato board will run rings around multiple servers for off-server write intensive tasks, such a compilations. One thing that I'm not sure of is if the Legato board will run with the drivers for the Interphase or the Ciprico. Larry Blair ames!vsi1!lmb lmb@vicom.com