On Tue, Sep 13, 2005 at 08:57:23PM -0400, michael chang wrote: > On 9/13/05, Brooks Davis <[EMAIL PROTECTED]> wrote: > > On Tue, Sep 13, 2005 at 08:15:56PM -0400, michael chang wrote: > > > On 9/13/05, Santanu Das <[EMAIL PROTECTED]> wrote: > > > > michael chang wrote: > > > > > > > > >Oh, I must have misread your statement. These dual-cpu nodes... what > > > > >kind of CPUs do they use? How many CPUs does the OS _think_ there > > > > >are? i.e. are they dual-core CPUs or something? > > > > > > > > > > > > > > Those are all 2.8 Xenon processors; are they dual-core CPUs??. I'm > > > > running SL3 and it thinks there are 4 CPUs on every node. > > > > > > I don't know about Xenon processors, but I do know that Intel Xeon > > > processors come in a Dual-Core _flavour_ (meaning that each "CPU" may > > > or may not have multiple "CPUs" in them (technically, they're "cores" > > > but anyways...)). > > > > > > AFAIK, In a Dual Core CPU, it still has to give each core a CPU ID or > > > otherwise I'm not sure how the OS is supposed to handle it. > > > > Dual-core cpus come with the same it set as Hyperthreaded CPU (so that > > licensing that was fixed for HTT continues to work), but they add > > another bit saying that they aren't actually virtual cores. An integer > > If I understand you correctly, this is meaning that they say that > they're real cores as opposed to virtual ones?
On dual core systems, yes, but to a first approximation they look like virtual cores because the CPU vendors want to trick existing software into not requiring more licenses since that's not what the user expects. > > model of CPU counts is extremely close to worthless these days. > > So do we count hostnames instead? Or IPs? Or what? What happens if a > PC has two hostnames and two interfaces and two ganglia instances... > although that'd be solved by eliminating the need to run two ganglia > instances. (I'm assuming that you'd need to run two of them now -- I > don't know for sure, I forget. Sorry.) > > Kinda gets confusing what we want to count... You either count something and admit it the count is fairly meaningless (that's our current approach) or you don't count at all and present an accurate model of the system which describes the hierarchy of resources and where they share things such as caches and busses. A simple count of physical sockets with things in them or virtual cores or real cores only tells part of the story. It's probably more useful then clock rate, but not by much. A simple count model is attractive because it's easy to delude your self into thinking it's accurate, but it's increasingly more common for it to misrepresent reality (just wait for 16-core, hyperthreaded CPUs :). -- Brooks -- Any statement of the form "X is the one, true Y" is FALSE. PGP fingerprint 655D 519C 26A7 82E7 2529 9BF0 5D8E 8BE9 F238 1AD4
pgpjK9Iik1fYn.pgp
Description: PGP signature
