amit@umn-cs.cs.umn.edu (Neta Amit) (05/30/88)
We are running a PC-LAN of some 20 nodes, with a gateway to the main departmental Ethernet. A dedicated AT with 3C501 card serves as a gateway and mail server (no, we don't use POP), as well as file server and print server for the IBM PC Network it's running. The Gateway software has been developed locally, and our ip/tcp is derived from MIT's and CMU's pc/ip. Mail was first installed over a year ago, Telnet is a few months old, FTP is in development. The machines are fairly loaded, but the load on the gateway is light. The gateway does as little a processing as possible, e.g. incoming packets are sent immediately to their destination, where they are assembled. As a result, the gateway's capacity is much better than that reported in similar projects. Last night we did a little experiemnt which might be of some interest to this newsgroup. We tried to overload the gateway, by connecting 14 nodes to it. Two nodes sent mail messages consisting of the file /etc/hosts (.25MB). The other 12 were logged in remotely, mostly to local machines, cat'ing /etc/hosts one by one, in a staggerring fashion. When we turned the debugging mechanism on to trace packets at the gateway, we noticed that remote sessions were running at a reasonable speed, where as Mail was flowing very slowly. When the 10th node started to cat, the gateway choked and died, due to a previously undetected bug. In part, this was due packet buffer being full. The buffer's capacity is 20 packets: 4 are permanently assigned to Mail, 2 in 2 out; another 2 are also dedicated; 14 are for telnet/ftp. The software is executing in the Small memory model, close to the limit, hence expansion of the buffer is not very likely. We then invoked a second gateway. We have the capability to statically assign different nodes to different gateways on the same subnet, with no overlapping. The second gateway is a standard AT node, NOT a file and print server of the PC-LAN. Reconfiguration requires rebooting. The improvement was instantaneous. Mail began to flow 8 times faster. Gateway buffers were only half-full. No problem has been noticed. --Neta Amit (amit@umn-cs.cs.umn.edu) University of Minnesota CSci -- Neta Amit U of Minnesota CSci Arpanet: amit@umn-cs.cs.umn.edu