The following message is a courtesy copy of an article that has been posted to bit.listserv.ibm-main,alt.folklore.computers as well.
shmuel+ibm-m...@patriot.net (Shmuel Metz , Seymour J.) writes: > PC's may have the speed edge for an individual I/O, but how does a fast PC > I/O stack up against 100's of concurrent mainframe I/O's? It's the number > of channels operating in parallel that gives the mainframe a speed edge. re: http://www.garlic.com/~lynn/2010f.html#7 What was the historical price of a P/390? http://www.garlic.com/~lynn/2010f.html#8 What was the historical price of a P/390? lot of it was adding lots of additional processors for managing the i/o programming. the 3033 using channel directors (158 engines with 370 microcode removed and only the integrated channels) wasn't all that hot. the 3090 had more processors for doing i/o. however, the 3880 disk controller had much slower processor for command processing that it significantly drove up channel busy time per operation. the result was that there was change in 3090 to add a lot more channels (for a "balanced configuration") to spread the 3880s across a larger number of channels. This pushed 3090 channel circuits passed a threshold and another TCM had to be added. POK wanted to charge the san jose disk division the cost of the extra TCM on every 3090 sold. 3090 was also being sold into some of the "supercomputer" market with vector processing. However, that also implied lots of real high-speed disks operating at HIPPI speeds (basically standards version of cray channel) operating at 100mbyte/sec. The 3090 i/o interface couldn't handle the 100mbyte/sec transfers ... so there was a hack to cut into the side of the extended store bus to added HIPPI. The problem there was no channel processors on the extended store bus ... just the 4k move instructions ... so 3090 HIPPI i/o had to be done with "peek/poke" paradgim. a lot of the sequent NUMA-Q machine was PC processors with lots of things like enormous amounts of i/o processing. old emails http://www.garlic.com/~lynn/2010e.html#email951030 http://www.garlic.com/~lynn/2010e.html#email961211 from recent post http://www.garlic.com/~lynn/2010e.html#70 Entry point for a Mainframe? about customer providing sequent with 3590 drives to get support done ... and mention that first pass of the sequent dynix 3590 device driver didn't support scatter/gather & sili. the application was several hundred million accounts with possibly tens of millions of transactions every day. the transactions would be sorted in account order (and account summary information on tape was in account sorted order). the application would read input tape, apply/merge transaction summary information with days transactions and write the result to new tape. the idea was to do the processing at full 3590 speed ... getting nightly processing done to approx 30mins elapsed time (compared to having every night processing taking a couple weeks elapsed time using various other approaches). a big bottleneck for mainframe has been the half-duplex channel paradigm and CKD simulation (attempting to compensate for the bottleneck results in significantly increased complexity). original harrier was dual 80mbit/sec links ... running asynchronously ... getting 160mbit/sec aggregate ... SSA doubled that to 160mbit/sec asynchronous links ... getting 320mit/sec aggregate (and running asynchrnously help offset increasing latency issue). As mentioned SSA was offered on s/390 integrated server: http://www-01.ibm.com/common/ssi/rep_ca/1/897/ENUS198-211/index.html mentioned here http://www.garlic.com/~lynn/2010f.html#8 What was the historical price of a P/390? compared to half-duplex 200mbit/sec escon ... which suffered protocol latency issues with half-duplex activity. FCS disk infrastructures ... operating similar to harrier/SSA but at FCS speeds ... long before FICON. that was sort of what got us into trouble (with mainframe group) ... with cluster scaleup ... referenced in this old post about jan92 meeting in ellison's conference room http://www.garlic.com/~lynn/95.html#13 and this old email references http://www.garlic.com/~lynn/lhwemail.html#medusa within a couple weeks after the meeting in ellison's conference room, the project was transferred ... announced as a product in the numerical intensive market (only) ... and we were told we couldn't work on anything with more four processors. Old press article from 17feb92 (approx. five weeks after meeting in ellison's conference room): http://www.garlic.com/~lynn/2001n.html#6000clusters1 and another from 19jun92 http://www.garlic.com/~lynn/2001n.html#6000clusters2 -- 42yrs virtualization experience (since Jan68), online at home since Mar1970 ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@bama.ua.edu with the message: GET IBM-MAIN INFO Search the archives at http://bama.ua.edu/archives/ibm-main.html