vBulletin Search Engine Optimization
| |||||||
| Register | FAQ | Members List | Calendar | Search | Today's Posts | Mark Forums Read |
| ||||
| I've got a puzzling CPU VP issue that I'm hoping someone else here has dealt with. I've got a dedicated database server running IDS 10.00.FC8 on a four-core RHEL Opteron server. We've configured four CPU VPs, one per core. There are almost always five to six threads sitting on the ready queue, despite the fact that the system is generally > 60% idle time. This sparks the age-old debate: add more CPU VPs than physical processor cores? Or is there something incorrectly configured about the _current_ setup that's causing my CPU VPs to do less than they should / yield too easily, etc.? |
| |||
| do you have set MULTIPROCESSOR 1 in $ONCONFIG Superboer. On 16 mei, 02:28, tgirsch <tgir...@NOSPAM.gmail.com> wrote: > I've got a puzzling CPU VP issue that I'm hoping someone else here has > dealt with. I've got a dedicated database server running IDS 10.00.FC8 > on a four-core RHEL Opteron server. We've configured four CPU VPs, one > per core. There are almost always five to six threads sitting on the > ready queue, despite the fact that the system is generally > 60% idle time. > > This sparks the age-old debate: add more CPU VPs than physical processor > cores? Or is there something incorrectly configured about the _current_ > setup that's causing my CPU VPs to do less than they should / yield too > easily, etc.? |
| |||
| tgirsch wrote: > I've got a puzzling CPU VP issue that I'm hoping someone else here has > dealt with. I've got a dedicated database server running IDS 10.00.FC8 > on a four-core RHEL Opteron server. We've configured four CPU VPs, one > per core. There are almost always five to six threads sitting on the > ready queue, despite the fact that the system is generally > 60% idle time. If the system is showing idle time, and you have threads sitting in the ready queue, then there is somthing which is preventing the cpuvp from getting scheduled by the OS. What is the wait IO time? What does top look like? > > This sparks the age-old debate: add more CPU VPs than physical processor > cores? Or is there something incorrectly configured about the _current_ > setup that's causing my CPU VPs to do less than they should / yield too > easily, etc.? |
| |||
| Madison Pruet wrote: > tgirsch wrote: >> I've got a puzzling CPU VP issue that I'm hoping someone else here has >> dealt with. I've got a dedicated database server running IDS >> 10.00.FC8 on a four-core RHEL Opteron server. We've configured four >> CPU VPs, one per core. There are almost always five to six threads >> sitting on the ready queue, despite the fact that the system is >> generally > 60% idle time. > > If the system is showing idle time, and you have threads sitting in the > ready queue, then there is somthing which is preventing the cpuvp from > getting scheduled by the OS. What is the wait IO time? What does top > look like? > > >> >> This sparks the age-old debate: add more CPU VPs than physical >> processor cores? Or is there something incorrectly configured about >> the _current_ setup that's causing my CPU VPs to do less than they >> should / yield too easily, etc.? Seen the same thing on HP when you get to larger number of procesors, larger IDS ready queues but low OS CPU usage. Put it down to a OS bug Cheers Paul |
| |||
| Paul Watson (Oninit) wrote: > Madison Pruet wrote: > >> tgirsch wrote: >> >>> I've got a puzzling CPU VP issue that I'm hoping someone else here >>> has dealt with. I've got a dedicated database server running IDS >>> 10.00.FC8 on a four-core RHEL Opteron server. We've configured four >>> CPU VPs, one per core. There are almost always five to six threads >>> sitting on the ready queue, despite the fact that the system is >>> generally > 60% idle time. >> >> >> If the system is showing idle time, and you have threads sitting in >> the ready queue, then there is somthing which is preventing the cpuvp >> from getting scheduled by the OS. What is the wait IO time? What >> does top look like? >> >> >>> >>> This sparks the age-old debate: add more CPU VPs than physical >>> processor cores? Or is there something incorrectly configured about >>> the _current_ setup that's causing my CPU VPs to do less than they >>> should / yield too easily, etc.? > > > Seen the same thing on HP when you get to larger number of procesors, > larger IDS ready queues but low OS CPU usage. Put it down to a OS bug > > Cheers > Paul RAM vs SWAP vs IDS memory? |
| |||
| [cutting] >>>> >>>> This sparks the age-old debate: add more CPU VPs than physical >>>> processor cores? Or is there something incorrectly configured about >>>> the _current_ setup that's causing my CPU VPs to do less than they >>>> should / yield too easily, etc.? >> >> >> Seen the same thing on HP when you get to larger number of procesors, >> larger IDS ready queues but low OS CPU usage. Put it down to a OS bug >> >> Cheers >> Paul > > RAM vs SWAP vs IDS memory? 48GB vs 48GB vs 12GB Cheers Paul |
| |||
| Try the following onconfig changes... VPCLASS cpu,num=4,noage,aff=0-3 This will turn on noage and affinity for your 4 cpu vps NETTYPE onsoctcp,1,600,NET FASTPOLL 1 Try using a NET VP instead of inline polling. This assume you do have some cpu bandwidth to allow the NET VP some cpu time. A single NET VP can handle up to 600 connections. If you have more than this, you might need a 2nd vp. On May 17, 9:45*am, "Paul Watson (Oninit)" <p...@oninit.com> wrote: > [cutting] > > > > >>>> This sparks the age-old debate: add more CPU VPs than physical > >>>> processor cores? *Or is there something incorrectly configured about > >>>> the _current_ setup that's causing my CPU VPs to do less than they > >>>> should / yield too easily, etc.? > > >> Seen the same thing on HP when you get to larger number of procesors, > >> larger IDS ready queues but low OS CPU usage. Put it down to a OS bug > > >> Cheers > >> Paul > > > RAM vs SWAP vs IDS memory? > > 48GB vs 48GB vs 12GB > > Cheers > Paul |
| |||
| pokeyman76@yahoo.com wrote: > VPCLASS cpu,num=4,noage,aff=0-3 > This will turn on noage and affinity for your 4 cpu vps > > NETTYPE onsoctcp,1,600,NET > FASTPOLL 1 I've seen stuff on FASTPOLL, but have no idea what it does. We're still considering turning on affinity. Madison Pruet wrote: > If the system is showing idle time, and you have threads sitting in the > ready queue, then there is somthing which is preventing the cpuvp from > getting scheduled by the OS. What is the wait IO time? What does top > look like? That's my understanding, but I can't find any evidence of this. As I said, idle time runs high (>60%), and wio is always < 1%. This is why I can't reconcile the ready queue with the high idle time. Although, to be fair, there are rarely more than about a dozen threads on the ready queue. Still, if idle time is that high, I wouldn't expect to see much of ANYTHING on the ready queue. |
| ||||
| tgirsch said: > pokeyman76@yahoo.com wrote: > > VPCLASS cpu,num=4,noage,aff=0-3 > > This will turn on noage and affinity for your 4 cpu vps > > > > NETTYPE onsoctcp,1,600,NET > > FASTPOLL 1 > > I've seen stuff on FASTPOLL, but have no idea what it does. We're still > considering turning on affinity. The executive summary, as explained to a slow clown was this: Current polling technology asks connection 1, "Have you got anything for me to do?" Current polling technology asks connection 2, "Have you got anything for me to do?" Current polling technology asks connection 3, "Have you got anything for me to do?" Current polling technology asks connection 4, "Have you got anything for me to do?" Current polling technology asks connection 5, "Have you got anything for me to do?" Current polling technology asks connection 6, "Have you got anything for me to do?" ... Current polling technology asks connection N, "Have you got anything for me to do?" Then it starts all over again. Whereas FASTPOLL asks: "Has anyone got anything for me to do? Has anyone got anything for me to do? Has anyone got anything for me to do? Has anyone got anything for me to do? Has anyone got anything for me to do? Has anyone got anything for me to do? Has anyone got anything for me to do? Has anyone got anything for me to do? Has anyone got anything for me to do? Has anyone got anything for me to do? Has anyone got anything for me to do? Has anyone got anything for me to do? ..." On HP, it apparently is like free money (because HP's polling libraries are shit.) On other platforms it's still worth doing, but not as impressive. -- Bye now, Obnoxio "There were a myriad of problems which conspired to corrupt your reason and rob you of your common sense. Fear got the best of you, and in your panic you turned to the Labour Party. They promised you order, they promised you peace, and all they demanded in return was your silent, obedient consent." |