Choose timezone
Your profile timezone:
Powered by Indico
v3.3.5-pre
DRAFT
Present: DP, Guilherme, Enric, Enrico Xavi, Gerri, Vassil
We need to have a topical meeting about vectorisation. Tentatively we will have it once Pere is back from holiday, mid July.
Enrico: Progress for 6.12
- Actions now are jitted only before the loop optimising runtime. It is not yet possible (needs thinking) to apply the same strategy for filters and defines. We can keep this in mind and perhaps react once this becomes a bottleneck.
- ROOT-8883: filling histos in parallel
* Either we save the entries individually and then we fill: memory issue
* We estimate the range and then we lock to broadcast the ranges
* We start from histograms which we extend coherently to then merge them at the end
* We can keep a buffer per thread and lock to fill a single histo
We need a decision. Gerri proposes that it might be fair to impose a slowdown in case the user does not want to specify anything.
- Early detection of missing branch names: working on 8879, 8873, 8880. Make this rock solid as this is a requirement for the data source. We need to be sure in a streamlined way to define a miss and go to the source.
Round table
- Enric: starting the benchmarks about ROOT-8862. The study is about the representation in memory of the cached data. We want to compare the time needed to read a file from disk, an uncompressed TMemfile, N vectors<T> where T is the type of the columns.
- Guilherme: ROOT-8871 mixed items for the improvement of performance of ROOT I/O. Full set of performance measurements and solutions to the single performance degradation identified. A suggestion by Danilo is to always doublecheck with a simple "time" measurement that the runtime reported by VTune is not heavily altered.
- Vassil: very complete benchmark written and published with a pr, 676. We agree that vectorisation should not be exposed in the very high level interfaces. Xavi can contribute for the presentation for intel the tests he prepared for his work about fitting.