Eviews 7 multicore efficiency on AMD

[Ruslan]

Hello.

Just bought Eviews 7.

Could you please tell to me is it any possibility to receive multicore efficiency on AMD (Phenom2 X4 955).

1 core load only (25% load overal). Any possibility how to increase total number of loaded cores and make 100% load?

Thank you.

[EViews Gareth]

EViews 7 should use the maximum number of cores you have automatically. You can check that it is set to do this by looking at Options->General Options->Advanced System Options->Multi-Processor/Multi-core use. It should be set at "Auto". Although you could always manually set it to the number of cores you have.

Note that EViews 7 won't switch to using all of the cores (i.e. 100%) unless you perform a task that would benefit from using multiple cores rather than a single core. It generally requires a large number of observations (1000s, or even 10,000s for some procedures) before it is beneficial to switch to using multiple cores. Also note that many time-series based calculations can not be performed on multiple cores at all.

[EViews Gareth]

Try running the following code. You should see the CPU pegged at 100% If it isn't, let us know!

Code: Select all

create u 1000000 group xs for !i=1 to 20 series X!i=nrnd xs.add x!i next series y=nrnd equation e1.ls y c xs

[Ruslan]

Yes! Total load achived.

For this structure:
-------------------------
create u 1000000
group xs
series y=nrnd
for !i=1 to 400
series X!i=nrnd
xs.add x!i
equation e1.ls y c xs
next
-------------------------

Equations load: Ls - 30-40%, VAR - 70-80% .

Adding MA(1) to LS - variable 25-30% total processor's load.
The same with ARCH(1,1) - 25% total processor's load. For the GARCH looks to be the same.

Thank you, anyway. Good job.

Best wishes,
Ruslan

[EViews Glenn]

To followup, Gareth's comments about many time series estimators not using multiple cores is relevant in the examples you cite. The problem is that estimators which require recursive evaluation can't be split up into sets of operations which can be passed off to different cores.

[Ruslan]

Sure.

It might be impossible to parallelise sequential estimation, where elements at step (t) depend on elements estimated at step (t-1). Roughly, the elements at step(t) is sent to core 1, thus, core 2 have to wait for the arrival of step(t) results, in order to start the estimation at step(t+1). So, the use of different cores in sequential estimation where each stage of estimation depends on previous stage gives nearly same effect as one core usage.

[EViews Glenn]

This analysis is correct and why sequential estimators cannot, in general, take advantage of multi-core processing.