sometimes being parallel may not be the right tradeoff in a memory constraint setting
kapalani
Performing compression before sending the data over to memory can improve the performance of a bandwidth limited program even though we perform more computation to uncompress the data when we read it back since the speedup of the program is limited by its usage of memory not arithmetic. However there is a tradeoff in that if we compress data too much then we spend more time decompressing it that it defeats the benefits of reading less from memory
jedi
Where do wimpy nodes fit in on this spectrum? Is there still a trend in moving away from more powerful compute nodes to clusters of cheap, energy-efficient cores, or was this never adopted widely (perhaps Intel did not want to be associated with "wimpy" things)?
sometimes being parallel may not be the right tradeoff in a memory constraint setting
Performing compression before sending the data over to memory can improve the performance of a bandwidth limited program even though we perform more computation to uncompress the data when we read it back since the speedup of the program is limited by its usage of memory not arithmetic. However there is a tradeoff in that if we compress data too much then we spend more time decompressing it that it defeats the benefits of reading less from memory
Where do wimpy nodes fit in on this spectrum? Is there still a trend in moving away from more powerful compute nodes to clusters of cheap, energy-efficient cores, or was this never adopted widely (perhaps Intel did not want to be associated with "wimpy" things)?