| PLEX86 | ||
Performance and Capacity Planning 734IIRC, the only reason JMF invented his spin lock is because KL caches were not write-thru and the other CPU had to wait for the data to get to memory. That's why our SMP has a cache sweep serial number. I may very well be confused here because all of this is based on my memory of conversations.
Would you give an example or two of your critical paths? It just occurred to me that these might be different because of OS philosophy differences. I'd always buttumed that these would be the same, no matter what was running on the hardware. Performance and Capacity Planning 738 re: planning planning it was interesting period at the science center concurrently and-or overlapped ... i was getting to do a bunch of ecps stuff all the invention... This was on a per-CPU basis, right? It sure seems like people made things difficult. Oh, then this microcode wasn't in each CPU....Interrupts had to global...or are you talking about CPU-specific interrupts like channels momentarily "owned" by a CPU? This sounds like the thingie JMF and TW called the doorbell, only upsidedown. A lot of our kernal was sharable. I can't remember the macro name but it surrounded the code that couldn't be sharable. It was a goal to have everything sharable. Sheesh. They must have been spending all their knocking each other up. My gut feel says that 25% is too much. We didn't have any control over this since we only implemented what hardware handed us. TOPS-10's approach was to make all that controller and device driver code reentrant. Then it doesn't matter which CPU uses what. A CPU can look over a queue of I-O requests and pick one that needed gear the CPU was physically connected to.
Performance and Capacity Planning 735 all vamps processors ran identical microcode and instructions. in vamps, the global kernel lock metaphor ... just precluded more than one processor at a time executing in... Yes, this is how it should be...completely event driven based on the usage of the system now. A lot of developers couldn't think this way; they were always insisting that our unknown users would be happen to squeeze their round jobs through our square holes.
Performance and Capacity Planning 736 the issue in cp67 was that all time was accounted for .... while in virtual machine mode ... it was all charged... YES! This way the sysadmin could add and remove CPUs as if they were another convenient device. Never have to take down the system to replace any gear. If a user needed a CPU that did special computing, s-he could place a mount request and the job could be scheduled after the super-duper adder was wheeled in from the lab next door. But then, I'm dating myself when I'm talking about wheeling gear in and out of rooms. ;-) BAH Subtract a hundred and four for e-mail.
|
||||
Performance and Capacity Planning 735 Alt Folklore Computers from Newsgroups The #1 Usenet Provider on the Internet
|
||||