cmcmanis%pepper@Sun.COM (Chuck McManis) (03/24/89)
In article <13392@steinmetz.ge.com> (Donald P Perley) writes: >I am not all that hot on diskless nodes myself. NFS is great, being able >to access files all over the place, but remember that you are limited to >a total of about 1.2 mbyte/sec. When you start splitting that among >a large number of stations, it would be nice to at least have memory >paging on a local disk. A couple of points about diskless nodes : 1) Diskless Amigas are 98% the speed of diskfull Amigas because once they are running something they don't swap. Loading and directory scanning is about the same as a non-DMA harddrive. 2) The ethernet has never been the bottleneck for diskless nodes. The order link strength is generally Server-CPU, Server-Disk Speed, Ethernet. At Sun we keep improving the Server's CPU and Disk Speed and have yet to get it to the point where either the server or the client was waiting to shove bits over the wire. The other point you missed (and don't feel bad, most people do) is that ethernet is not "split" among the various diskless nodes, it is "given" to them. Thus whenever you need it, you get _all_ 1.2 MBytes/Sec of bandwidth and don't have to share with anyone. No one makes a disk drive for the Amiga yet that can pump out 1.2 Mbytes a second. (Yes they come close, but still they aren't there yet.) --Chuck McManis uucp: {anywhere}!sun!cmcmanis BIX: cmcmanis ARPAnet: cmcmanis@sun.com These opinions are my own and no one elses, but you knew that didn't you.
perley@trub.steinmetz (Donald P Perley) (03/25/89)
In article <95626@sun.Eng.Sun.COM> cmcmanis@sun.UUCP (Chuck McManis) writes: >In article <13392@steinmetz.ge.com> (Donald P Perley) writes: >>I am not all that hot on diskless nodes myself. NFS is great, being able >>to access files all over the place, but remember that you are limited to >>a total of about 1.2 mbyte/sec. When you start splitting that among >>a large number of stations, it would be nice to at least have memory >>paging on a local disk. > >A couple of points about diskless nodes : > > 1) Diskless Amigas are 98% the speed of diskfull Amigas because > once they are running something they don't swap. Loading and > directory scanning is about the same as a non-DMA harddrive. I was refering to all the brew-ha-ha about Commodore needing to supply their own ethernet card to have a legitimate unix workstation, and be able to boot diskless. Doesn't AMIX have paging? > > 2) The ethernet has never been the bottleneck for diskless nodes. > The order link strength is generally Server-CPU, Server-Disk Speed, > Ethernet. At Sun we keep improving the Server's CPU and Disk Speed > and have yet to get it to the point where either the server or > the client was waiting to shove bits over the wire. Each of the resources (server-cpu, server-disk, ethernet) are a shared resource. We probably have close to 500 diskless nodes here, and dozens of servers. If all of a servers clients are clamouring for packets at the same time, things can get kind of bogged down. A resource is easier to share if only one person wants it at a time. The thing about your paging file is that your machine is the only one that needs it, so there is no need to use a shared resource to handle it. (unless the per megabyte cost is a lot less on the server) >The other point you missed (and don't feel bad, most people do) is that >ethernet is not "split" among the various diskless nodes, it is "given" >to them. Thus whenever you need it, you get _all_ 1.2 MBytes/Sec of bandwidth >and don't have to share with anyone. No one makes a disk drive for the >Amiga yet that can pump out 1.2 Mbytes a second. (Yes they come close, >but still they aren't there yet.) Maybe I was mistaken about the nature of ethernet. I thought you only get it all if no one else is asking. If a server tries to send a packet can't it get a collision and have to retry? -don perley
cmcmanis%pepper@Sun.COM (Chuck McManis) (03/26/89)
In article <13425@steinmetz.ge.com> (Donald P Perley) writes: >Each of the resources (server-cpu, server-disk, ethernet) are a shared >resource. We probably have close to 500 diskless nodes here, and >dozens of servers. If all of a servers clients are clamouring for >packets at the same time, things can get kind of bogged down. 500 diskless clients and servers on one ethernet is non optimal its true. I would hope you would break those into several IP nets with an occasional server doubling as a gateway (extra ethernet cards are about the same price as a Mac level bridge but _much_ more effective at balancing the network load.) >A resource is easier to share if only one person wants it at a time. Also true. >The thing about your paging file is that your machine is the only one that >needs it, so there is no need to use a shared resource to handle it. >(unless the per megabyte cost is a lot less on the server) This is also true. Lots of people here have been playing with the idea of the "swap" disk. A small, high speed SCSI disk (like the Quantum 40S) that has no file system on it, just swap space. >>The other point you missed (and don't feel bad, most people do) is that >>ethernet is not "split" among the various diskless nodes, it is "given" >>to them. Thus whenever you need it, you get _all_ 1.2 MBytes/Sec of bandwidth >>and don't have to share with anyone. No one makes a disk drive for the >>Amiga yet that can pump out 1.2 Mbytes a second. (Yes they come close, >>but still they aren't there yet.) >Maybe I was mistaken about the nature of ethernet. I thought you only >get it all if no one else is asking. If a server tries to send a packet >can't it get a collision and have to retry? Your mistaken in how ethernet actually works in real life. (Which is different than you would expect it to by reading a description of CD/CSMA type networks) The two key points are that usually the clients and servers are in "sync" which means that the client will use the net to send a packet, and then wait for a reply that the server will send. And that the machine first checks the wire, if there are bits out there it waits until it is clear, otherwise it blasts out a packet. Only if someone else picked that *exact* same time (to within a few dozen microseconds) will there be a collision. The effect in reality is that each participant "sees" a 1.2MByte/s hose that they can use to talk to the other person with. The client sends a request over a 1.2Mb line, and the server answers. Now in the diskless case what actually tends to happen is the the clients will all send a request for a page or what ever and then all go dormant waiting for the server to respond. The server sorts out all the requests and queues them all out to the clients. I agree it is non-intuitive but this is really how it works. --Chuck McManis uucp: {anywhere}!sun!cmcmanis BIX: cmcmanis ARPAnet: cmcmanis@sun.com These opinions are my own and no one elses, but you knew that didn't you. "A most excellent barbarian ... Ghengis Kahn!"