[email protected] (David Crayford) writes: > But is that unique to a mainframe? When you compare it to a POWER > system or whatever Oracle are flogging these days it doesn't stand > out. Even commodity servers hooked up to an enterprise class HBA can > handle massive amounts of I/O throughput.
re: http://www.garlic.com/~lynn/2013m.html#94 SHARE Blog: News Flash: The Mainframe (Still) Isn't Dead http://www.garlic.com/~lynn/2013m.html#96 SHARE Blog: News Flash: The Mainframe (Still) Isn't Dead it possibly was 3090 when IBM marketing respun the requirement for significant increase in the number of channels. fundamentally, mainframe channel is considered busy when it is doing lots of end-to-end back&forth channel programming protocol chatter. original number of 3090 channels were originally done based on assumption that 3880 disk controller was as efficient as the (previous) 3830 disk controller. However, the 3880 disk controller had a much slower processor for handling channel protocol activity ... which enormously increased channel busy. when POK finally realized the significance, they had to significantly increase the number of channels in order compensate for the enormous increase in channel busy (as a result of the slower 3880 processor) as a means of achieving target I/O throughput. The increase in channels resulted in needing an additional TCM ... which increases the 3090 manufacturing cost. At the time, there was semi-facetious suggestion that POK bill the 3880 disk controller group for the increase in 3090 manufacturing costs. In any case, marketing then respins the significant increase in 3090 channels (from needing to compensate for significant increase in channel busy by 3880) to the significant amount of aggregate i/o bandwidth (even if it couldn't be effectively used). harkens back to 70s during the future system period when marketing really got its reputation for FUD compensating for lack of competitive products (the lack of new products during the FS period is credited with giving clone processors a market foothold). Note that part of 3090 (and vector processor facility) was targeted for supercomputer market. However important part of supercomputer market was high-speed i/o, 100mbyte/sec (800mbit/sec) "HiPPI" channel (standardization work for cray channel out of LANL) ... for things like large disk arrays (for instance thinking machines had a 32+8 disk array ... i.e. parallel transfer across 32 drives). http://en.wikipedia.org/wiki/HIPPI from above: To understand why HIPPI is no longer used, consider that Ultra3 SCSI offers rates of 320 MB/s, and is available at almost any corner computer store. Meanwhile Fibre Channel offered simple interconnect with both HIPPI and SCSI (it can run both protocols) and speeds of up to 400 MB/s on fibre and 100 MB/s on a single pair of twisted pair copper wires. ... snip ... standard 3090 i/o interface was totally unable to handle 100mbyte/sec operation. What some Kingston engineers did was cut into the side of the extended store memory bus and implement a peak/poke paradigm for HiPPI i/o ... the 4k-byte extended bus synchronous move instructions were used to create i/o commands at reserved extended store memory locations. as part of ha/cmp cluster scaleup work ... we got periodically dragged into both LANL and LLNL (as well as other national labs). serial scsi http://en.wikipedia.org/wiki/Serial_attached_SCSI serial ata http://en.wikipedia.org/wiki/Serial_ATA as previously mentioned, I had hoped that 9333/harrier would have evolved into interoperable fractional FCS ... as opposed to non-interoperable S http://en.wikipedia.org/wiki/Serial_Storage_Architecture -- virtualization experience starting Jan1968, online at home since Mar1970 ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [email protected] with the message: INFO IBM-MAIN
