Re: architectural goals, Byte Addressability And Beyond

Lawrence D'Oliveiro <ldo@nz.invalid> writes:

Remember that “mainframe workloads” are primarily I/O bound, not CPU-
bound. The whole concept of a “mainframe” arose in the era when CPU time
was scarce and expensive, so you had all these intelligent I/O peripherals
that could be given sequences of operations to perform, with minimal CPU
intervention. It was all about maximizing throughput (batch operation),
not minimizing latency (interactive operation).

1980, I was con'ed into helping IBM STL lab that was overcrowded and
moving 300 people to offsite bldg with dataprocessing back to STL
datacenter. I was asked to do channel-extender support to place "local"
channel "attached" controllers at the remote bldg (cutting various
protocol round-trip latencies). Part of the issue was that the mainframe
60s era also had limited memory and so there is enormous protocol
round-trips utilizing data back in mainframe memory.

1988, local IBM branch asks if I could help LLNL national lab
standardize some serial stuff they had been playing ... which quickly
becomes fibre-channel standard (FCS). Some time later, some IBM
engineers become involved with FCS and define a heavy-weight protocol
that radically cuts the native throughput ... which was eventually
released as FICON (used for mainframe I/O, w/extensive protocol
round-trip latencies, significant impact for even short distrances at
gbit rates).

Most recent public benchmark that I've found is IBM z196 "Peak I/O"
benchmark which had 104 FICON (running over 104 FCS) getting 2M IOPS.
About the same time, a native FCS was announced for E5-2600 blade
claiming over million IOPS (two such FCS have higher throughput than
104 FICON (running over 104 FCS). Note IBM docs recommend that SAPs
(CPUs dedicated for running I/O) be kept to no more than 70% CPU
... which would be more like 1.5M (rather than 2M) IOPS.

--
virtualization experience starting Jan1968, online at home since Mar1970