In trying to figure out the best number of wu's to run on any given machine I've always used the APR as a rough guideline for estimation of overall system throughput. If you run one wu, the number is higher, but I have always assumed that if you are running 2 or 3 work units at a time you at the reported APR and multiply it by the number of instances you are running on the GPU. I've never seen anyone say that explicitly but it seemed to make sense. Is my assumption correct?

Thanks,

Chris

---

While APR may help, it is also worth keeping an eye on the execution times, which give you a more immediate view of what is going on.

---

Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

In trying to figure out the best number of wu's to run on any given machine I've always used the APR as a rough guideline for estimation of overall system throughput. If you run one wu, the number is higher, but I have always assumed that if you are running 2 or 3 work units at a time you at the reported APR and multiply it by the number of instances you are running on the GPU. I've never seen anyone say that explicitly but it seemed to make sense. Is my assumption correct?

Thanks,

Chris

While APR does, indeed, decline when you run multiple tasks per GPU, I don't know that you can apply a direct multiplier such as you suggest and come up with a meaningful result. As rob smith recommended, manually comparing run times (at different angle ranges), is likely to be much more accurate.

APR can actually be fairly volatile, as I found when I tried to use it to make some app selection decisions of my own a few years ago. Take a look at a thread that I started back then, titled What, exactly, does APR measure, and is it meaningful?. In particular, I think you might find the very knowledgeable replies of Richard, Claggy, and Joe Segur to be quite informative.

---