hyperthreading CPU benchmark joke

I tried this benchmark suite called INQUISITOR on an
old Pentium 4 (with HT).
I runs Dhrystones, Whetstones test on core 0 then core 1 !
WTF?
It's the same bloody CPU.
The programmer must have got his diploma out of a cornflakes box.

On 13/02/2011 7:51 AM, Orson Cart wrote:
I tried this benchmark suite called INQUISITOR on an
old Pentium 4 (with HT).
I runs Dhrystones, Whetstones test on core 0 then core 1 !
WTF?
It's the same bloody CPU.
The programmer must have got his diploma out of a cornflakes box.


The method for detecting a true multi-core vs. a virtual multi-core is
exactly the same, involving using the CPUID instruction in a specific
way. So the difference to software of whether it's a true core or a
Hyperthreading core is nil.

Besides, the differences between virtual cores and true cores is
starting to blur these days. For example, AMD has come up with the
concept of the "Bulldozer" cores which are somewhat more than a virtual
core, but somewhat less than a full core; they say two cores in a
Bulldozer module should be about 80% as powerful as a full core, but
sharing all caches with each other and using much less energy. So the
benchmark may actually end up being the appropriate way to look at cores.

Yousuf Khan

On Feb 17, 8:11*am, Yousuf Khan wrote:
On 13/02/2011 7:51 AM, Orson Cart wrote:

* I tried this benchmark suite called INQUISITOR on an
old Pentium 4 (with HT).
I runs Dhrystones, Whetstones test on core 0 then core 1 !
WTF?
It's the same bloody CPU.
The programmer must have got his diploma out of a cornflakes box.

The method for detecting a true multi-core vs. a virtual multi-core is
exactly the same, involving using the CPUID instruction in a specific
way. So the difference to software of whether it's a true core or a
Hyperthreading core is nil.


Not if you are scheduling. You want a scheduler to be aware of
whether a given processor is a distinct physical core or not. Linux
SMT Kernel scheduling has been "hyperthreading-aware" for a few years
at least. How it is actually done is above my pay grade.

Robert.

On 17/02/2011 8:44 PM, Robert Myers wrote:
Not if you are scheduling. You want a scheduler to be aware of
whether a given processor is a distinct physical core or not. Linux
SMT Kernel scheduling has been "hyperthreading-aware" for a few years
at least. How it is actually done is above my pay grade.

Robert.

There is a way to figure it out through the same CPUID instruction. I
think it involves looking for bits in the return register's upper vs.
lower nibble or something like that.

Yousuf Khan

Robert Myers wrote:
On Feb 17, 8:11 am, Yousuf wrote:
On 13/02/2011 7:51 AM, Orson Cart wrote:

I tried this benchmark suite called INQUISITOR on an
old Pentium 4 (with HT).
I runs Dhrystones, Whetstones test on core 0 then core 1 !
WTF?
It's the same bloody CPU.
The programmer must have got his diploma out of a cornflakes box.

The method for detecting a true multi-core vs. a virtual multi-core is
exactly the same, involving using the CPUID instruction in a specific
way. So the difference to software of whether it's a true core or a
Hyperthreading core is nil.


Not if you are scheduling. You want a scheduler to be aware of
whether a given processor is a distinct physical core or not. Linux
SMT Kernel scheduling has been "hyperthreading-aware" for a few years
at least. How it is actually done is above my pay grade.

You are correct, the problem is you have to know if HT is faster or slower than
using two cores. If you are running pthreads in the same process, working on
data in memory, sharing an L1 cache may be faster than two full cores. For other
loads the opposite is true.

I'm not disagreeing, just saying knowing HT from cores isn't the sum of
information needed for best scheduling.