TINS: A Task-Based Dynamic Helper Core Strategy for In Situ Analytics

Abstract

The in situ paradigm proposes to co-locate simulation and analytics on the same compute node to analyze data while still resident in the compute node memory, hence reducing the need for postprocessing methods. A standard approach that proved efficient for sharing resources on each node consists in running the analytics processes on a set of dedicated cores, called helper cores, to isolate them from the simulation processes. Simulation and analytics thus run concurrently with limited interference. In this paper we show that the performance can be improved through a dynamic helper core strategy. We rely on a work stealing scheduler to implement TINS, a task-based in situ framework with an on-demand analytics isolation. The helper cores are dedicated to analytics only when analytics tasks are available. Otherwise the helper cores join the other cores for processing simulation tasks. TINS relies on the Intel R TBB library. Experiments on up to 14,336 cores run a set of representative analytics parallelized with TBB coupled with the hybrid MPI+TBB ExaStamp molecular dynamics code. TINS shows up to 40% performance improvement over various other approaches including the standard helper core.

Modeling high-throughput applications for in situ analytics

“…As usually there is a small number of analysis functions, this exponential factor remains limited (cf. table 1 in the study by Dirand et al (2018) for an example).…”

“…Damaris (Dorier et al, 2012), FlowVR (Dreher and Raffin, 2014), Functional Partitioning (Li et al, 2010), GePSeA (Singh et al, 2009), Active Buffer (Ma et al, 2006), or SMART (Wang et al, 2015) adopt this solution. Tins (Dirand et al, 2018) introduced dynamic helper cores, dedicating cores to analysis only when analysis tasks are ready to be run. Even if the in situ processing simply consists in saving data to disks, this approach can be more efficient than to rely on standard I/O libraries like MPI I/O (Dorier et al, 2012).…”

“…Analytics tasks can include about anything, like computation of high-level descriptors (Dirand et al, 2018), compressing data (Tao et al, 2017), verifying the integrity of the data to detect a silent error (Aupy et al, 2013) or checkpointing for reliability (Aupy et al, 2014).…”

“…GoldRush proposes to activate analysis executions only on long-enough sequential sections of the simulation. Tins adopts a task-based programming relying on a work-stealing scheduler for a fine grain interleaving and load balancing of tasks on the compute nodes (Dirand et al, 2018).…”

DEISA: Dask-Enabled In Situ Analytics

“…Task-based programming where the tasks are dynamically distributed to compute resources is today classical for shared memory programming using for instance OpenMP or Intel TBB. TINS [21] leverages this approach as long as the simulation is also parallelized on each node with tasks. TINS relies on the TBB work-stealing scheduler to dynamically distribute the tasks on the cores, being simulation or analytics tasks.…”