Preparation for benchmark in hpc

[georgegito]

No description provided.

[georgegito]

For increasing number of processes, average nodes each query point visits is increasing, but total time is decreasing:

• Parallelism outruns this visits increase.

• Check cache misses.

• Analyze VPT search behavior for increasing number of processors.

[georgegito]

Need to declare a reasonable b value with respect to local n.

#Update: b = 0.3 * log2(n / world_size)

[Stavrosfil]

We can also measure how long it takes to calculate each local knn based on the tree, and how long each data transaction takes, to search for possible bottlenecks.

[georgegito]

v1: Almost perfect speedup (check cache misses).
v2: The number of visited nodes increases by increasing processors, so speedup is a little worse.

[georgegito]

We can also measure how long it takes to calculate each local knn based on the tree, and how long each data transaction takes, to search for possible bottlenecks.

Sure, we could choose some of our datasets and analyze them in depth.

[Stavrosfil]

I will perform an in depth cache miss and branch miss analysis for the single threaded instance of the problem, and maybe a simple profiling report for the speedup analysis.
Appropriate plots will be included in the final report.

[georgegito]

We could compare v1 versus v2 average communication time and analyze the tree sending time.

[Stavrosfil]

Yes, each possible procedure in the algorithm should be analyzed to an extend. Communication times, tree build and rebuild and processing times should be measured for comparisons.

We should also take b and y block size into account, both for speed and cache misses.

[georgegito]

Yes, each possible procedure in the algorithm should be analyzed to an extend. Communication times, tree build and rebuild and processing times should be measured for comparisons.

We should also take b and y block size into account, both for speed and cache misses.

If we take b and y block size into account too, we will need 10 pages for the report..

[Stavrosfil]

Let's move the discussion to #23 to avoid clutter here.