Which Kernel to use, .EL or .ELsmp

Lew Bloch conrad at lewscanon.com
Sat May 28 22:58:36 UTC 2005


Quoting Gavin McDonald:

>We run dual-Xeons w/ Hyperthreading on an IBM eServer, and with the smp
>kernel we get (the appearance of?) 4 CPUs.
>
>Actually, this makes me wonder...  Am I truly getting 4 threads running
>synchronously?  Maybe I should watch 'top' output more closely, and see if
>the individual loads are paired,
>or evenly spread...
>
>Any thoughts, all?
>
Intel's HT (HyperThread) technology is a clever hack that makes the P4 a 
very nice CPU out of an inefficiency in the original design.  In the 
original P4 the prefetch queues were so deep that they caused execution 
stalls, and it was actually slower than a similarly-clocked P3 would 
be.  Then Intel added circuitry to run two independent execution threads 
simultaneously.  The stalls on one side give execution time to the other 
CPU thread; they just become opportunities for parallelism.  Now one P4 
chip acts in every respect like two, and the increase in speed is 
comparable to a true two-CPU brew.  Of course, the use of only one 
physical chip means a simpler motherboard with less power consumption.  
So a P4 liability became a huge asset.

 From the OS's point of view, a HT CPU is two CPUs.  The hyper-threading 
is done electrically, in the chip itself, so it presents in every 
respect as two physical CPUs.   It is not a mere appearance of multiple 
CPUs.

As for how well Linux (or any other OS) uses those extra CPUs, that's 
another question.  If it uses actual multiple CPUs well it'll use the HT 
P4 well, and contrariwise.




More information about the redhat-list mailing list