> The zAAP may not have a need to use the 360-MP "shoulder tap" technique to > start I/Os or to field I/O interrupts. (I haven't found an appropriate > dump yet - I have other things to do now.) But I wouldn't actually > suspect it uses that antique technique since (a) zAAPs are limited to > processors using PR/SM and (b) Chris, whom I trust, said they don't use > it. PR/SM definitely treats the zAAPs special so there may well be sneaky > hipervisor techniques used to signal the processors... especially if the > processors needing to be signalled are cross-book, for example.
Thanks for the vote of confidence Tom, but I don't think that's really what I said. I don't know exactly how dispatching is done on the zAAP. >From conversations with BCP folks and knowing a bit about the innards of the OS, I can speculate how it might be done, but that's all. To the best of my knowledge the actual workings are undisclosed. > I haven't even established whether zAAPs run enabled or disabled for I/O > interrupts yet... if they ran disabled for I/O they would, of course, run > a lot faster. I have not actually looked at a zAAP to be sure, but I am almost certain they do run disabled for interrupts. That isn't so odd, since most of the engines in a multi-engine LPAR are actually disabled for interrupts most of the time. Think CPENABLE. What's actually going on under the covers in that case is that the OS disables interrupts via the control register, so the state of the I/O interrupt mask in the PSW is irrelevant. The PSW mask has to be and-ed with the control register, so no interrupts are actually seen by that engine, even though it nominally appears to be enabled for interrupts. The system selectively enables more engines for fielding I/O interrupts based on the number of interrupts handled via TPI. It's a throughput versus responsiveness trade-off. Given the limitations of zAAPs there would be no reason to enable one for I/O interrupts. > (All of the zAAP stuff seems like a shameful waste of creative programming > by Greg D. & Co. just to keep the ISVs from draining the customer base > dry.) Pretty much. Greg, Peter Relson and others burned a lot of cycles doing it. My question at the time was... couldn't you just have implemented separate accounting fields for JVM work? Apparently lawyers were involved in making that decision, so this is what we get for allowing lawyers to dictate system design. CC ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [EMAIL PROTECTED] with the message: GET IBM-MAIN INFO Search the archives at http://bama.ua.edu/archives/ibm-main.html
