Hypercomm::Aggregation Library

Introduction

This is the aggregation/aggregator component of the Hypercomm suite of libraries for Charm++. It supports virtual routing, and chare-arrays, chare-groups, and chare-node-groups.

Usage

To use the library, declare an initproc calling aggregation::initialize within your .ci file:

namespace aggregation {
	initproc void initialize(void);
}

Then, include hypercomm/aggregation.hpp within your C++ code.

We recommend using a std::unique_ptr to hold your aggregator, declaring it as such:

std::unique_ptr<aggregator_t> aggregator;
...
aggregator = std::unique_ptr<aggregator_t>(new aggregator_t(
	/* message threshold, e.g. 100 */,
	/* flush timeout, e.g. 0.1s */,
	[this](void*  msg) { /* endpoint functionality */ },
	/* node level, true/false */
	/* ccd periodic condition, e.g. 10ms, idle, etc. */));
);

Send calls should go through the aggregator itself, e.g.:

aggregator.send(idx, ...);
proxy->ckLocalBranch()->aggregator->send(idx, ...);

In node-level scenarios, the aggregator is thread-safe and does not need an external lock 😊

Usage Notes

All PEs should initialize all their aggregators in the same order, creating one instance for each endpoint per PE (e.g., ep1 then ep2 for all PEs). This is to ensure their sequential IDs are assigned consistently. Note, in the future, we may add an option for users to specify an endpoint ID (eliminating this constraint). We recommend using a node/group or Ckpv variables to hold aggregators, then accessing them via the ckLocalBranch or CkpvAccess respectively.

Compile-Time Options

Hypercomm has a handful of compile-time configuration options:

• HYPERCOMM_NODE_AWARE — A boolean flag (0 or 1) that indicates whether the aggregator should be node-aware in SMP scenarios.
• HYPERCOMM_TRACING_ON — Enables Hypercomm's internal tracing routines, this generates various end-of-run summary statistics.