On Fri, 24 Jun 2005, Clayton Macleod wrote: > actually, top's "load average" has very little to do with CPU usage. > All it tells you is how many processes were competing for / using *any > amount* of CPU in a given time period. It doesn't tell you what > percentage of CPU is being used, you can have a load average in the > 5.x range and still have the CPU sitting idle 80% of the time. Load > average does help you guage performance, but it's only one factor of > many that tell you how the box is actually performing and whether or > not it is underpowered for the tasks it is doing. If the first of the > three load average numbers is 5.2, for argument's sake, all that tells > you is that in the last minute there was an average of 5.2 processes > each using *some* CPU time at once.
You are contradicting yourself or you have a scew picture of what load is. You are first talking about "how many processes were competing" and then you write "all that tells you is that in the last minute there was an average of 5.2 processes each using *some* CPU time at once." Which is not the same thing and is not really the truth. The correct desription of load is more like: Load: An average of how many processes that where in queue for getting kernel resources. Have a look at http://www.hostpronto.com/article/36 for more info on UNIX load. > And I'm afraid everyone here that has said HT skews the CPU usage > tracking are entirely correct, you can't accurately tell how much CPU > is actually being used and how much is actually free/idle when you > have HT turned on. It's just a fact of HT life. It doesn't relate at > all to a true dual CPU machine when it comes to measuring how much CPU > is being used and how much is idle, because it's difficult to actually > measure when you're only using tricks to emulate dual CPUs. > HT definitely helps out *some* in a multi-threaded/multi-tasking > environment, but its very nature means you can't really tell how much > it will help out. Nor can you easily measure exactly how much idle > time you have left, because of the slight of hand that's involved. But > if you turn HT off you can very accurately tell how much CPU is > actually being used. So if you're having performance problems, and you > need to know how much CPU is really being used, that is indeed a very > good step to take. Turning HT off and running the same scenario again. This is also not really true. You will get a good look how much each virtual processor uses. At least in FreeBSD. In FreeBSD 5.x you can do a simple 'ps aux | grep idle' and get how much idle each virtual processor has. Also all the %CPU usage in top and ps for a process only shows on a virtual CPU, so in theory you can have 4 processes using 99.999% CPU without hitting the roof. But I agree, he should try to disable HT. In theory you will get an performance boost, since the SMP will now only need to have to know about 2 instead of 4 CPU's. You also is not bound to half the CPU power for the process. So if a process needs to have more than what a HT CPU can offer, it will then get that without hitting the roof. The only reason I can see why you should have HT enabled is that if one process hangs and start using up 100% of CPU, you still have 75% of the total cpu power left to work with instead of 50%. But if that happens on a gameserver, you probably jump to the server directly to fix it anyway. There is also a bug in the HT that allows in certain circumstances to read data from other processes. Due to this bug, FreeBSD now have HT disabled by default. You need to switch it on with a sysctl variable. If I could I would have switched of HT on all my servers. But due to a bug in FreeBSD, it will not boot up. I get some strange IDE timeouts at bootup. Perhaps upgrading from 5.3 to 5.4 will solve it. /Bjorn -- Sometimes...in the morning...when I haven't slept...the computers talk to me, they say "good morning dave", so I say to myself, "That g**D**N Hal's awake!" _______________________________________________ To unsubscribe, edit your list preferences, or view the list archives, please visit: http://list.valvesoftware.com/mailman/listinfo/hlds_linux
