glen herrmannsfeldt <[email protected]> writes: > For unix/dos/windows file systems, though, there needs to be a disk > cache where FBA blocks are brought into memory and the appropriate > bytes copied to user space. Now, the caching ability of that likely > helps much of the time, but it isn't so different from what CKD > emulation has to do.
re: http://www.garlic.com/~lynn/2012m.html#2 Blades versus z was Re: Turn Off Another Light - Univ. of Tennessee http://www.garlic.com/~lynn/2012m.html#3 Blades versus z was Re: Turn Off Another Light - Univ. of Tennessee but major server apps & rdbms on those platforms do direct disk i/o w/o system caching (and/or rdbms do their own caching) .... they also do direct i/o for network lan. the application direct disk i/o was part of what drove POSIX asynch I/O specifications in the late 80s. problem is since there haven't been industry standard benchmarks published for mainframe ... it takes some digging trying to uncover apples-to-apples. FICON architecture layered half-duplex channel paradigm on top of industry standard fibre channel. As a result, it needed to keep track of "open exchanges" ... here it increases max from 60 to 600 ftp://ftp.software.ibm.com/common/ssi/sa/wh/n/zsw03059usen/ZSW03059USEN.PDF above also has z10 ficon express4 31,000 maximum 4kbyte I/Os per second or 124MBYTE/sec base fibre channel didn't need that overhead ... just send out i/o program to controller/device for execution. this has zhpf at 92,000 4k channel i/os per second and 368MBYTE/sec increases to 1600MBYTE/sec with large sequential read/write mix http://www-03.ibm.com/systems/z/hardware/connectivity/ficon_performance.html this has z/os max. z196 with peak of 2million 4k I/O ops/second ftp://public.dhe.ibm.com/common/ssi/ecm/en/zsw03169usen/ZSW03169USEN.PDF from above: zHPF improves upon FICON by providing a Transport Control Word (TCW) that facilities processing of an I/O request by the channel and the contorl unit (and improves throughput compared to original FICON CCW at a time processing). The TCW has capability that enables multiple channel commands to be sent to the control unit as a single entity instead of as separate comands as in FICON CCW (zHPF/TCW partially implements the original underlying fibre channel design point from the late 80s nearly 35yrs ago). Figure 6 on pg9 shows how zHPF improves over FICON ... and starts to approach the throughput of the native fibre channel implementation and the original native fibre channel design point from 35yrs ago. The 2M 4k I/O ops/second is max 80 processor z196 with (max) 14 system assist processors (theoritically maximum is 2.2M SSCH running all 14 system assist processors at 100% utilization) and 104 FICON Express8 channels to 11 storage subsystems. this claims for e5-2600, latest emulex delivers over one million IOPS on a single channel and doubles the bandwidth of previous generation http://www.emulex.com/artifacts/0c1f55d0-aec6-4c37-bc42-7765d5d7a70e/elx_wp_all_hba_romley.pdf One question are the FICON Express8 channels actually 104 different fibre-channel channels? ... aka the 2.2M ops are spread across 104 different fibre-channel channels ... while a single Emulex fiber-channel channel is able to do 1M ops. The previous reference has LSI storage subsystem capable of 724K IOPS and peak 5Gbyte/sec while the Adaptec storage subsystem is 450K IOPS and 6.6Gbyte/sec The mainframe numbers appear that it can do peak of 92k IOPS on single channel with zHPF (& TCW with some i/o program batching attempting to approach base fibre channel paradigm) and maxed. out z196 is capable of 2M IOPS with 14 dedicated system assist processors and 104 channels (although theoritically 104*92 = 9.6M IOPS ... while each emulex channel is capable of 1M IOPS). -- 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
