At work recently, I found myself trying to explain the Work-Efficiency vs Step-Efficiency tradeoff to a coworker, but when I searched for online resources to help I couldn’t find any that I liked, so I decided to take a shot at writing my own. I found this idea presented in a video lecture series about programming for GPGPUs on Youtube a while ago. However, it’s just as applicable to any form of parallel processing, from SIMD instructions running on a single CPU core up to massive clusters of thousands of computers.
Good explanation of the difference between work efficiency and step efficiency when talking about parallel algorithms.
Yes, I was thinking about that some more. Synchronization can be large in terms of time if you have work units that do not have known time or count. Actually this is an exception too. That is you may want to have number of work units large compared to number of cores for example. Which I think is actually opposite from what you said except of course the more work units probably the more total overhead.
except of course the more work units probably the more total overhead.
Yeah this is exactly it: when you start splitting work between multiple processors, there’s often more overhead involved in simply doing that splitting and coordinating the results. That’s why an explicitly non-parallel algorithm’s work efficiency is going to be as good or better than a parallel one. It’s impossible to parallelize an algorithm and make it more work efficient than the non-parallel version, even though it’ll (hopefully 😅) use less wall clock time due to having better step efficiency, because it doesn’t have to do the whole thing sequentially.
Yes. In the end it is usually total wall time of the critical path that matters unless there are special considerations like power consumption or heat generation, etc.
Not the same thing, but CPUs are designed to optimize different things too. I know I have often had laptops that were faster for single threaded tasks, but my engineering workstation has typically had way higher throughput for a parallel work load. Say nothing about compute clusters. Lot of these things about single threaded and parallel fairly common.
There also tends to be fairly expensive setup and sync steps related to highly parallel stuff like MPI or GPUs, though sometimes less so in the case of say vectorization or threading.
Yes. Took my first computer class in 1978. Fortran of all things. Hooked immediately. Considered CS, but actually went the Science side of things doing a combination of design, simulation, instrument control, and machine control. Whatever was needed. Lot of changes in hardware, software, languages, etc. on one hand but on the other hand not so much. At some level it has all been the same.
Yes, I was thinking about that some more. Synchronization can be large in terms of time if you have work units that do not have known time or count. Actually this is an exception too. That is you may want to have number of work units large compared to number of cores for example. Which I think is actually opposite from what you said except of course the more work units probably the more total overhead.
Yeah this is exactly it: when you start splitting work between multiple processors, there’s often more overhead involved in simply doing that splitting and coordinating the results. That’s why an explicitly non-parallel algorithm’s work efficiency is going to be as good or better than a parallel one. It’s impossible to parallelize an algorithm and make it more work efficient than the non-parallel version, even though it’ll (hopefully 😅) use less wall clock time due to having better step efficiency, because it doesn’t have to do the whole thing sequentially.
Yes. In the end it is usually total wall time of the critical path that matters unless there are special considerations like power consumption or heat generation, etc.
Not the same thing, but CPUs are designed to optimize different things too. I know I have often had laptops that were faster for single threaded tasks, but my engineering workstation has typically had way higher throughput for a parallel work load. Say nothing about compute clusters. Lot of these things about single threaded and parallel fairly common.
There also tends to be fairly expensive setup and sync steps related to highly parallel stuff like MPI or GPUs, though sometimes less so in the case of say vectorization or threading.
Thanks, interesting.
It is interesting, isn’t it? This is why I love computing as a field; the whole “spectrum” from theory to practice is just super fascinating.
Yes. Took my first computer class in 1978. Fortran of all things. Hooked immediately. Considered CS, but actually went the Science side of things doing a combination of design, simulation, instrument control, and machine control. Whatever was needed. Lot of changes in hardware, software, languages, etc. on one hand but on the other hand not so much. At some level it has all been the same.