home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!hela.iti.org!usc!isi.edu!finn
- From: finn@isi.edu (Greg Finn)
- Newsgroups: comp.dcom.cell-relay
- Subject: Re: Computers dont like ATM?
- Message-ID: <22909@venera.isi.edu>
- Date: 18 Nov 92 19:38:06 GMT
- References: <1992Nov18.025754.14749@trl.oz.au>
- Sender: news@isi.edu
- Reply-To: finn@dalek.isi.edu (Greg Finn)
- Organization: USC-Information Sciences Institute
- Lines: 54
-
- In article <1992Nov18.025754.14749@trl.oz.au> you write:
- >This morning I heard a talk by Dr Paul Green from IBM Research ...
- >More or less in passing, he mentioned that 53 byte cells are too small
- >for computers (or words to that effect). His claim was that small
- >packets ... make a lot of unnecessary work for the computer that has to
- >receive them, since it has to process a header every 53 bytes ...
- >He also seemed to be suggesting that this storage and retrieval process
- >could become a bottleneck (at 100s of Mbit/s?) because of the limitations
- >of memory bandwidth.
-
- Very true. I co-developed a multi-gigabit LAN. If a
- store-and-forward (or retrieve) operation is to occur without reducing
- the channel bandwidth, it requires memory that operates at twice the
- channel rate. Our simplex channels operate nominally at 500 Mb/s and
- require 33 MHz memory. Using CMOS it is very difficult to create bulk
- memory that operates at 33 MHz, let alone 66 MHz.
-
- Packet storage within a switch should be avoided at high
- channel rates if practical and can be accomplished with cut-through
- routing. This issue has nothing whatsoever to do with fixed versus
- variable-length packets. Avoiding the need for storage within a
- switch is even more important in the all-optical domain.
-
- >He was advocating variable length packets (an idea
- >called PTM - packet transfer mode - that IBM has proposed), because it
- >puts the complexity in the switch rather than in the receiving
- >terminal.
-
- I know nothing about the desirability of PTM in telco
- switching. I can say something about LANs. PTM allows one to design
- LANs that are exceedingly fast and that have tiny host interfaces that
- require no fragmentation/reassembly hardware. PTM allows one to make
- LAN switches as simple as ATM-based LAN switches if the right routing
- technology is chosen. Source routing and a variant of cut-through now
- used in multicomputers provided us a datagram routing decision of
- about 30ns per hop and yields a 10 Gb/s 32x32 crossbar that is smaller
- that an 8.5x11 inch sheet of paper.
-
- >From this, his conclusion was (I think) that ATM is good for switching
- >voice traffic, but not so good for computer networks. This seems to go
- >against what's actually happening, with people building ATM LANs while
- >telcos hang back, not sure if ATM is the right way to go.
-
- ATM is politically very popular. Arguments that ATM brings
- about end-to-end compatibility do not impress me. Most LAN traffic
- stays within the LAN. Forcing LANs to comply with WAN data-link
- standards is putting the cart before the horse. I have great
- reservations about what ATM costs us in LANs. When ATM-based LAN
- performance exceeds PTM-based LAN performance/cost, ATM proponents
- will have a much better argument than they now have. Frankly, I don't
- forsee that happening in the electrical domain.
- --
- Gregory Finn (310) 822-1511
- Information Sciences Institute, Marina Del Rey, CA 90292
-