E-Locus - Institutional Repository of the University of Crete

Home Collections School/Department School of Sciences and Engineering Department of Computer Science Post-graduate theses

Post-graduate theses

Current Record: 52 of 824

[Add to Basket]

Identifier

000449175

Title

An efficient and lightweight OpenSHMEM implementation

Alternative Title

Μία αποδοτική και ελαφριά OpenSHMEM υλοποίηση”

Author

Καλυβιανάκης, Γεώργιος Δ.

Thesis advisor

Πρατικάκης, Πολύβιος
Καλλιμάνης, Νικόλαος

Reviewer

Μαγκούτης, Κωνσταντίνος
Δημακόπουλος, Βασίλειος

Abstract

High Performance Computing (HPC) is rapidly gaining momentum, relying on the benefits of the Partitioned Global Address Space (PGAS) model for optimal results. Numerous languages and libraries have been introduced that leverage the PGAS model, with the most widely known being OpenSHMEM. OpenSHMEM is a standard specification that defines a onesided, RDMA capable API for extensive use in HPC. In this thesis we design and implement Gmem, an OpenSHMEM implementation supporting TCP/IP, RoCE and Infiniband networking protocols, backed by GSAS, a very lightweight PGAS API, allowing processes spawning on a number of nodes to communicate in very similar way to shared memory schemantics. Gmem leverages shared memory for intranode communications, enabling users to fully utilize spacial locality without involvement of the OS or the network adapter. With RDMA we are also able to perform operations on remote nodes with extremely low latency and high throughput. We evaluate Gmem with the OpenSHMEM implementation of OpenMPI and MPICH that rely on the Unified Communication X (UCX) framework, for TCP/IP and Infiniband. In our tests we assess the performance of PUT/GET remote memory operations, several atomic memory operations and collectives operations. During our experiments, we observe that our implementation is not only on par with our competitors, but in some cases we even achieve greater results. In GET operations, for large size transfers we achieve 6x lower latency than OpenMPI, and in Atomic operations 1.25x better latency than OpenMPI.

Language

English

Subject

Distributed systems

HPC

Infiniband

Issue date

2022-07-29

Collection

School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses

Type of Work--Post-graduate theses

Permanent Link