On 16/11/2012 2:16 AM, Anne & Lynn Wheeler wrote:
IBM has z196 benchmark with peak of 2m IOPS with 104 FICON channels, 14 storage subsystems, and 14 system assist processors. It mentions that the 14 SAPs are capable of peak 2.2m SSCH/sec running at 100% cpu busy, but recommends SAPs run at 70% or less (i.e. 1.5m SSCH/sec). there is also a recent emulex announcement single fibre-channel for e5-2600 capable of over one millions IOPS (compared to z196 peak of 2m IOPS using 104 FICON channels)
It's difficult for most mainframers to fathom commodity servers capable of doing that kind of I/O. In the Xeon E5 architecture Intel created a new technology called DDIO which substantially reduces I/O latencies so all of those dual/quad socket servers just chomp it all up without waiting to be fed. Throw in an emulex HBA and you're cooking with gas! The speeds and feeds are mighty impressive, the first 16GBs PCIe-3 implementation to hit the market. When you consider that the high-end POWER boxes and mainframes only got PCIe-2 last year it's interesting to see how it will evolve. One things for sure, the squatty boxes can do the I/O and with virtualization they may seem attractive to a lot of companies for a whole plethora of workloads. It's a clear and present danger to the high-end RISC boxes. It's still hard to imagine mainframe customers jumping ship anytime soon. New applications like high frequency trading will almost certainly be deployed on x86 platforms.
---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [email protected] with the message: INFO IBM-MAIN
