Path: utzoo!news-server.csri.toronto.edu!cs.utexas.edu!uunet!math.fu-berlin.de!uniol!unido!materna!elwood!tb From: tb@Materna.DE (Torsten Beyer) Newsgroups: comp.sys.novell Subject: Re: NFS Support in NetWare Message-ID: Date: 15 Mar 91 11:56:15 GMT References: <1991Mar10.041835.6883@novell.com> Sender: root@Materna.DE Lines: 78 Keith, keith@ca.excelan.com (Keith Brown) writes: >one out. I believe you, but I'm having a little trouble finding out anything >about how they implemented it (true dynamic disk cache allocation). Perhaps I'm not sure, but I believe you won't find anything about that in the docs. Sun changed their memory management dramatically when they went to 4.x. The new design basically maps every fileacces to paging. So files don't get read in the way they used to in the old days (via the buffer cache) instead they get PAGED in. The trick now is that all pages reside on a page list which more or less works like an LRU list. So once a file get's paged in, it'll stay in it's pages until the kernel decides somebody else could use these pages in a better way. As all memory (except for kernel memory and static kernel structures) is in the page pool, this memory CAN act as a buffer cache. Considering a configuration where a system only acts as an NFS server, it's very likely that most unused processes will be swapped to disk, thus freeing up real memory as buffer pages. This is basicaly the way it works (if I remember correctly). I used to know all this a lot better, when I had access to the SunOS 4.0 source, which is about a year ago. Sun presented some papers about their memory management strategies at USENIX conferences (i think in 88 and 89). Offhand I don't know the exacy conferences and names of the authors. If you are interested I'll scan my desk and homedirectory to find more precise information. >1) At what point does SunOS say to itself "Hmm.. Nobody appears to be > interested in using me as a proper UNIX system. Seems all they want me > for is NFS serving. I guess I'll grab all this memory I've been sitting > on and stuff it in the disk cache." And Like I said, there is no explicit point at which a server says that. The "decision" depends on what the system does. Acting only as an NFS server will put the system in the described state within a minute or so. It's just that more and more pages are used by files. >2) What happens when a UNIX based (sorry, SunOS based) NFS server that > has made this desision is slumbering in it's machine room at 3 O'Clock > in the morning when I, heavily camoflaged as a 9 track tape drive, sneak > up to its console and give it a nasty surprise by logging in and firing > up the biggest make(1) its ever seen in its life? This is where your anti-UNIX argument is right. At this point the kernel will, according to it's paging strategy, throw out (file)pages in order to get memory for the make and it's children. This will certainly degrade the fileserving performance >And I'll certainly be pleased to show it to you. I'm at the mercy of our >German office regarding the equipment I'll have but hopefully they'll >provide me with something meaty enough to make it impressive. If I arrive Unfortunately this articel arrived here one day too late. I visited CeBIT yesterday and tried to find somebody at the Novell booth to show me your NFS stuff. They were however extremely unfriendly and told me it was not possible and nobody was demonstrating it... :-(. I'm certain somebody did, but I couldn't manage to find the right person (the boot was extremely packed). >In my situation, no! I'm not monitoring a nuclear reactor, niether am >I scanning the skies over Iraq looking for incoming scuds. If I was I'd >turn write through back on. I'm comfortable in the knowledge that my >NetWare NFS server is doing its damndest to flush my data to disk every >three seconds if it can. Besides, in my experiance, my client is far >more likely to crash than my server. Then where's all my valuable data? Hmm, we have a Novell-Net here and it basically works ok. Recently however we had problems with our power supply and the server would crash regularly (of course not it's fault). This led to me loosing one hours work of typing... This wouldn't have happened if I were using NFS. ciao -Torsten -- Torsten Beyer e-mail : tb@Materna.DE Dr. Materna GmbH VOX : +49 231 5599 225 Vosskuhle 37 FAX : +49 231 5599 100 D-4600 Dortmund 1, West Germany