Jim Sibley writes: > > > I'm curious. One of the benefits touted, and true, > about > > Linux on zSeries > > > vs. some other platform, is the zSeries' strength > in I/O. > > Is this still true > > > with FCP attached SCSI DASD? Why would the zSeries > drive > > SCSI DASD better > > > than Intel or Sun? > > > John McKown > > > Senior Systems Programmer > > Basically you can attach more dasd space and have more > simultaneous (NOT just concurrent) data transfers > going on at the same time. > > The I/O advantage of the mainframe is that it usually > has more paths (256 channels) to more devices(65,536) > thus giving a lot more parallel I/O, not that any > particular device is more efficient. If you have a lot > threads active, more I/O can be done in parallel that > most intel and other boxes. > > With 256 channels at say 12 MB/sec (shark) on , the > total aggregate rate of the mainframe would be about 3 > GB/sec. Obviously, that's limited by the 2 GB backend > buss on the TREXX.
The general idea is right but the bus limit is wrong: 2GB/sec I/O for an entire box would be very poor. Rather than have zSeries damned with faint praise, allow me to hype up its I/O capabilities a bit more. 2GByte/sec is the speed of a single STI bus and the smallest T-Rex (one book) has 12 STI buses while the largest (four books) has 48 STI buses for a total of 96GByte/sec bandwidth. Channel cards, whether ESCON or FICON, are spread over domains/slots to take advantage of the STI buses available. You can't fill all of that bandwidth with DASD I/O (there's a limit of 120 x FICON 2Gbit/sec ports--60 features on z990--making a nominal 24Gbyte/sec) but it's way more than 2GB. ESCON hits the limit of number of channels way before any hardware bandwidth limit but even so you only have 16 ESCON ports per card. Each STI bus fans out to four slots and, for ordinary I/O, gets multiplexed down to 333MByte/s, 500MByte/sec or 1000MByte/sec as appropriate. For ESCON, it uses 333Mbyte/s (which nicely encompasses the 16 x 20MByte/s nominal signalling for an ESCON card) and for FICON, 500MByte/sec (which nicely encompasses the 2 x 200MByte/s nominal for the dial-port FICON-Express cards). The buses and features are, IMHO, very well designed to ensure that there are no bottlenecks or caps right through to the backend memory bus adapters (MBAs) of the memory subsystem. For those interested in the details, Chapter 3 of the "z990 Technical Guide" redbook (SG24-6947) from www.redbooks.ibm.com elaborates on this and describes it very well. > Also, the main frame typically has 2 processors > dedicated to driving the devices (SAPs), so less "real > cpu" is used for I/O. In fact, not just the SAPs (which deal with initiating the I/Os). Each channel card also is fairly powerful and has the responsibility of doing much of the I/O work itself. For example, each z900 FICON card has two 333MHz PowerPC processors (cross-checked for reliability) to do the work. Again, for lots of detail, see the "z900 I/O subsystem" paper by Stigliani et al in the z900 edition of the IBM Journal of R&D (Vol 46 No 4/5 Jul/Sept 2002). --Malcolm -- Malcolm Beattie <[EMAIL PROTECTED]> Linux Technical Consultant IBM EMEA Enterprise Server Group... ...from home, speaking only for myself
