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: 50 of 824

[Add to Basket]

Identifier

000447836

Title

Design and implementation of a write based version of the Exanet MPI

Alternative Title

Σχεδιασμός και υλοποίηση μιας έκδοσης του Exanest MPI βασισμένης σε εγγραφή

Author

Νικολουδάκης, Μιχαήλ Β.

Thesis advisor

Πρατικάκης, Πολύβιος

Reviewer

Μαγκούτης, Κωνσταντίνος
Παπαευσταθίου, Βασίλειος

Abstract

MPI is one of the leading communication protocols used in HPC (High Performance Computing) suites today due to its portability and scalability. Many HPC applications make use of MPI in order to enable communication between different processes. In the scope of the ExaNeST project, an HPC prototype was deployed in the CARV Laboratory of FORTH consisting of 512 ARMv8 cores coupled with FPGA logic. This prototype makes use of special network primitives designed to allow the low latency transmission of control messages as well as the efficient transfer of large data through the Exanet network. In order to exploit the aforementioned capabilities of the prototype, a highly optimized MPI implementation (Exanet MPI) was developed in the scope of the same project prior to our work. This implementation makes use of the prototype’s communication primitives and manages to outperform the well known MPI implementation, MPICH by achieving up to 30x lower latency. Exanet MPI supports both an eager and a long communication protocol used for short and large MPI transfers respectively. The long protocol depends on emulated DMA reads and supports exclusively sender initiation. Sender initiation is defined as the ability of the sender of an MPI message to initiate the communication with the receiver by issuing an appropriate control message. Despite its simplicity, sender initiation does not let us exploit scenarios in which the receiver posts its request earlier than the sender. In addition, the use of emulated reads requires the receiver to notify the sender about the end of a DMA transfer through the use of an Ack control message which incurs extra latency. In this thesis, we design and implement from scratch a write-based version of the Exanet MPI that supports both sender and receiver initiation. With the use of DMA writes, we render the sender able to determine the end of a DMA transfer by itself without the need of acknowledgment from the receiver. Additionally, we take advantage of cases where a receive request gets posted earlier than a matching send request by letting the receiver initiate communication by asynchronously transferring its DMA related information to the sender. Consequently, a sender that posts its send request after the receiver, can immediately transfer data without the need of further synchronization with the receiver. However, simply adding receiver initiation support to the long protocol also infers some complications including (but not limited to) the significant increase of the eager protocol's latency. We propose our method for successfully facing the complications that arise from the support of receiver initiation and we also further optimize the long protocol by eliminating the need of some control messages. In order to break down the performance gain caused by our optimizations we develop in total 4 variants of the write based Exanet MPI. In each variant, we provide implementations for most point-to-point, collective as well as communicator manipulating functions. We describe the use cases of each developed variant and evaluate them against the already optimized read based original version of Exanet MPI on the HPC prototype. We offer insight into the ways our control path optimizations improve performance and the factors that let our implementation show more benefit. For the evaluation we use both microbenchmarks and real scientific applications. We show that our implementation can outperform the read based protocol by up to 50% in communication latency while also reduce the total execution time of specific applications by up to 10% (depending on the percentage of communication time they contain).

Language

English

Subject

Communication protocols

Message passing interface

Receiver initation

Διεπαφή μεταβίβασης μηνυμάτων

Πρωτόκολλα επικοινωνίας

Issue date

2022-07-29

Collection

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