Verdandi


Also found in: Dictionary, Thesaurus, Wikipedia.
Related to Verdandi: Verthandi

Verdandi

Norn of time present. [Norse Myth.: Wheeler, 260]
See: Time
References in periodicals archive ?
Therefore, in the case of Verdandi it is possible that synchronization issues will affect weak scalability even more seriously for future models, which will have higher connectivity.
From the plots it appears that pNeo is two to three times faster than Verdandi for the same problem size even without considering threads (otherwise it would be about 20 times faster) or the larger number of connections included in pNeo for larger network sizes.
Verdandi shows a flattening of the expected computation time for larger sizes--at least in part because the average number of synapses per neuron is becoming constant.
For Verdandi, to understand the different performance on the two architectures (about 4 times faster on Fusion on a per core basis), the effect of using a shared memory approach on computation time needs to be considered.
Verdandi can make very effective use of shared memory: for the simulations we tested changing from 1 to 24 threads increased the total memory used by only 25%.
In this case, the good fit we have with a linear model suggests that there is a sizeable serial part in the calculations, consistent with Verdandi's implementation and Amdahl's law [34], with about a 1/3 of the single thread time becoming serial, thus increasing the total time linearly in the number of cores because they would all sit idly, plus a 2/3 parallel part whose total time would be unaffected by the number of cores.
Now, let us reconsider Verdandi's performance difference in completing the same simulation as a hybrid calculation (MPI + OpenMP) on Beagle or as a pure MPI calculation on Fusion.
Verdandi spent a considerable amount of time on barriers, both for MPI (synchronizing communication) and OpenMP (mostly waiting for serial parts and synchronizing threads): the combined effect could easily put the total amount of idle CPU time over 90% of the total used for computation.
The memory utilization in large-scale simulations comparable to ours [7] was 2.8 TB for a 22 M neuron simulation having 11B synapses (about half of the connections expected from Verdandi for a network of this size, as Verdandi is already in a linear regime at 100 K, and much less than the number of connections expected in a pNeo simulation of that size).
Moreover, Verdandi did not include long-range interactions.