vbc...@gmail.com (Vince Coen) writes: > I think the stats on migration failures show that many fail regardless > of the target migration mainly is that they over estimate project > time, and quality of the target systems being used in place of m/f. > > Taking a straight view the mainframe is slow compared to running on > servers on a instruction throughput basis. > > What they miss however is the data through put specs compared to > mainframes where the m/f still wins hands down. > > I have tried (just for my self) to build a 8 core PC with separate > Sata controllers for each 15000 rpm drive to match up with m/f > performance but apart from the high costs of each controller there is > still the speed or lack of it of going from the controllers to the > application because of bottle necks in the data bus. > > I have not seen any PC/server design mobo that gets around this > problem and until they do - the mainframe is still "the man" for data > processing in bulk.
Lots of migration failures are trying to make any change at all. A simple scenario is the financial industry spent billions of dollars in the 90s to move from "aging" overnight (mainframe) batch settlement to straight-through processing using large numbers of parallel "killer micros". A major source of failure was wide-spread use of industry parallelization libraries (that had 100 times the overhead of cobal batch). I pointed it out at the time, but was completely ignored ... the toy demos looked so neat. It wasn't until they tried to deploy that they ran into the scaleup problems (the 100 times parallelization overhead total swamped the antificapated throughput increases using large number of "killer micros" for straight-through processing). In the meantime there has been enormous amount of work by the industry (including IBM) on RDBMS parallizing efficiencies. A RDBMS-based straight-through processing implementation done more recently easily demonstation all of the original objectives from the 90s ... but the financial industry claimed that it would be at least be another decade before they were ready to try again (lots of executives still bore the scars from the 90s failures and had become risk adverse). In 2009, non-mainframe IBM was touting some of these RDBMS parallelization scaleup efficienices. I somewhat ridiculed them ... "From The Annals of Release No Software Before Its Time" ... since I had been working on it 20yrs earlier (and got shutdown, being told I was not allowed to work on anything with more than four processors). Also, in 1980 I got sucked into to do channel extender for STL that was moving 300 people from the IMS group to off-site bldg. The channel extender work did lots of optimization to eliminate the enormous channel protocol chatter latency over the extended link ... resulting in no appearant difference between local and remote operation. The vendor then tried to get IBM approval for release of my support ... but there was group in POK working on some serial stuff (and were afraid if it was in the market, it would make releasing their stuff more difficult) and managed to get approval blocked. Their stuff is final released a decade later, when it is already obsolete (as ESCON with ES/9000). some past posts http://www.garlic.com/~lynn/submisc.html#channel.extender In 1988, I was asked to help LLNL standardize some serial stuff they have, which quickly morphs into fibre-channel standard (including lots of stuff that I had done from 1980). Later some of the POK engineers define a heavy weight protocol for fibre-channel that drastically reduces the native throughput which is eventually released as FICON. some past posts http://www.garlic.com/~lynn/submisc.html#ficon The latest published numbers I have from IBM is peak I/O benchmark for z196 that used 104 FICON (running over 104 fibre-channel) to get 2M IOPS. At the same time there was a fibre-channel announced for e5-2600 blade that claimed over million IOPS (two such fibre-channel has greater native throughput than 104 FICON running over 104 fibre-channel). In addition, there hasn't been any real CKD manufactured for decades, CKD is simulated on industry standard fixed-block disks. It is possible to have high-performance server blades running native fibre-channel with native fixed-block disks that eliminates the enormous FICON and CKD simulation inefficiencies. Related z196 I/O throughput number is all 14 SAPs running at 100% busy peaks at 2.2M SSCH/sec ... however, they recommend that SAPs are limited to 75% or 1.5M SSCH/sec. I have yet to see equivalent numbers published for EC12 or z13. EC12 press has been that z196 @ 50BIPS processing to EC12 @ 75BIPS processing (50% more processing) only claims 30% more I/O throughput. z13 quote has been 30% more processing than EC12 (with 40% more processors than EC12). Note that while fibre-channel wasn't originally designed for mainframe ... but for non-mainframe server configurations (that tend to run a few thousand), SATA is design point for the $500-$800 PCs. The are a lot of throughput differences between the consumer $500-$800 PCs and the non-mainframe server blades that are done for more heavy duty processing. -- virtualization experience starting Jan1968, online at home since Mar1970 ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN