Speaker
Alexey Rybalchenko
(GSI - Helmholtzzentrum fur Schwerionenforschung GmbH (DE))
Description
After Long Shutdown 2, the upgraded ALICE detector at the LHC will produce more than a terabyte of data per second. The data, constituted from a continuous un-triggered stream data, have to be distributed from about 250 First Level Processor nodes (FLPs) to O(1000) Event Processing Nodes (EPNs). Each FLP receives a small subset of the detector data that is chopped in sub-timeframes. One EPN needs all the fragments from the 250 FLPs to build a full timeframe. An algorithm is being implemented on the FLPs with the aim of optimizing the usage of the network connecting the FLPs and EPNs. The algorithm minimizes contention when several FLPs are sending to the same EPN. An adequate traffic shaping is implemented by delaying the sending time of each FLP by a unique offset. The payloads are stored in a buffer large enough to accommodate the delay provoked by the maximum number of FLPs. As the buffers are queued for sending, the FLPs can operate with the highest efficiency. Using the time information embedded in the data any further FLP synchronization can be avoided. Moreover, “zero-copy” and multipart messages of ZeroMQ are used to create full timeframes on the EPNs without the overhead of copying the payloads. The concept and the performance measurement of the implementation on a computing cluster are presented.
Primary author
Alexey Rybalchenko
(GSI - Helmholtzzentrum fur Schwerionenforschung GmbH (DE))
