home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!gatech!paladin.american.edu!auvm!UCLAMVS.BITNET!CSP1DWD
- Message-ID: <IBMTCP-L%92122319104540@PUCC.PRINCETON.EDU>
- Newsgroups: bit.listserv.ibmtcp-l
- Date: Wed, 23 Dec 1992 16:08:00 PST
- Sender: IBM TCP/IP List <IBMTCP-L@PUCC.BITNET>
- From: "Denis DeLaRoca 825-4580 (310)" <CSP1DWD@UCLAMVS.BITNET>
- Subject: Re: IBM TCP/IP inadequacies (was Usage of vector facility for
- TCP/IP?)
- Lines: 85
-
- Frank Vance <fvance@WG.WAII.COM> writes:
-
- > I used to think that, but Bill R. convinced me they really do care, y
- > just don't have the resources. So they try to focus on things they
- > think are most in demand from the customers.
-
- What resources? It is a fact that the TCPIP product is a proven money-
- maker, resources can be created easily to support such products. What
- triggered this and other messages is the fact that severe and well
- known problems in fundamental areas of the product (ie, IUCV/VMCF)
- persist and that IBM has publicly (thru this list and thru support)
- expressed their priority to fix these problems, well over a year
- later all these stated intentions and promises have vanished into
- thin air.
-
- You might ask why is IUCV/VMCF so important? As it turns out IUCV/VMCF
- (and the VM platform) make up the essential bridge between user
- processes and the TCPIP kernel, every single TCPIP transaction goes
- thru that bridge, but this bridge is neither solid nor ample, you can
- fall from it very easily. For example you can't reliable run one tcp/ip
- application in one split of my ISPF session and then start another
- tcp/ip app (whether it be from IBM or not) in a second split... I can't
- have that same tcp/ip app multitask and use tcp/ip services without the
- risk of losing access to IUCV from within my address space. Can't even
- run IBM's own telnet and FTP clients together... can't deploy Gopher,
- Wais, Web, NewsReader clients, or any other tcp/ip applications that
- multitasks and/or shells out to other tcp/ip applications without
- running into some TCPIP problem, usually an IUCV/VMCF problem. We have
- observed complex crashes in IUCV cleanup, we have observed user address
- spaces being left non-swappable by IUCV, the list goes on...
-
- The starting point for TCPIP was the VM work done at Wisconsin, but
- what was a virtuous, simple and glorified hack, in those days, has
- become totally obsolete and regressive. The current TCPIP architecture,
- unchanged from those simple days, bears absolutely no resemblance
- to modern day reality of fast computers and high-speed networking,
- the fact that TCPIP for MVS is hosted by means of a badly engineered
- (and buggy) VM emulator platform is absolutely mind boggling... the
- basic engineering principles to do high performance tcp/ip implemen-
- tations are by now well known, the sound engineering principles to
- host such implementations efficiently in modern day operating systems
- are also well known. MVS itself is in many respects as good an operating
- system as you can get and I am told by the MVS gurus that no significant
- obstacle exists to host a high-performance implementation of tcp/ip.
-
- It is clear that market pressures pushed IBM for an early release of
- TCPIP in what I'd consider a fundamentally flawed implementation, an
- implementation with inherent functional, performance and reliability
- limits. Rather than following that up with a better engineered imple-
- mentation we've seen more and more construction on a building that is
- for all intensive purposes sitting on a sandpit... and the crux of the
- matter is dealing with Operating System issues. To customer complains
- of poor performance we've seen the emergence of "offloading" whereby
- the basic tcp/ip protocol processing is offloaded to a peripheral
- processor... but this hardly changes the picture as we are still left
- saddled with complex interfaces to talk to an offload tcpip box, and
- we are still saddled with all the overhead and problems of the VM
- emulation platform. I don't see any future on this, on paper it
- looks as, if not more, expensive that the current solution.
-
- > My personal opinion is that the "typical" customer wants some crazy ngs,
- > like CICS sockets, the NDB system, and the like, none of which I hav
- > use for. Of course, that "typical" customer probably thinks my
- > requirement list (higher throughput, lower CPU consumption, better
- > Network Systems support) is strange too.
-
- I think these last requirements: performance, low resource utilization
- and rational network systems support are essential, nothing strange
- about them, and I would contend should be the areas that IBM should
- be putting all their resources and expertise to provide. I don't much
- care if they give me NFS and X-Windows support, I want the absolutely
- best tcp/ip engine that MVS is capable of hosting, with that building
- block in place the rest can be build around.
-
- But if today, right this minute you want a mainframe tcp/ip that can
- go 700-800Mb/s easily without overwhelming your CPU you go to the Cray
- folks not IBM. This at a time when gigabit networking and applications
- are just over the horizon doesn't augur too well for the future of
- tcp/ip on IBM mainframes. Remember, part of this discussion was tri-
- gerred by the guy trying to do some real networking and trying to
- diminish the ensuing 90% overhead on his computer!
-
- Something radical has to happen, what will it be?
-
- -- Denis
-