The Case for Transport-Level Encryption in Datacenter Networks
Authors:
Tianyi Gao,
Xinshu Ma,
Suhas Narreddy,
Eugenio Luo,
Steven W. D. Chien,
Michio Honda
Abstract:
Cloud applications need network data encryption to isolate from other tenants and protect their data from potential eavesdroppers in the network infrastructure. This paper presents SDP, a protocol design for emerging datacenter transport protocols, such as pHost, NDP, and Homa, to integrate data encryption with the use of existing NIC offloading of cryptographic operations designed for TLS over TC…
▽ More
Cloud applications need network data encryption to isolate from other tenants and protect their data from potential eavesdroppers in the network infrastructure. This paper presents SDP, a protocol design for emerging datacenter transport protocols, such as pHost, NDP, and Homa, to integrate data encryption with the use of existing NIC offloading of cryptographic operations designed for TLS over TCP. Therefore, SDP could enable a deployment path of new transport protocols in datacenters without giving up hardware offloading support, which would otherwise make encryption on those protocols even slower than TLS over TCP. SDP is based on Homa, and outperforms TLS over TCP by up to 29 % in throughput. SDP currently supports two real-world applications, Redis, improving throughput by up to 24 %, and in-kernel NVMe-oF, cutting P99 latency by up to 21 %.
△ Less
Submitted 21 June, 2024;
originally announced June 2024.
iFast: Host-Side Logging for Scientific Applications
Authors:
Steven W. D. Chien,
Kento Sato,
Artur Podobas,
Niclas Jansson,
Stefano Markidis,
Michio Honda
Abstract:
We have seen an increase in the heterogeneity of storage technologies potentially available to scientific applications, such as burst buffers, managed cloud parallel file systems (PFS), and object stores. However, those applications cannot easily utilize those technologies, because they are designed for traditional HPC systems that offer very high remote storage and network bandwidth. We present i…
▽ More
We have seen an increase in the heterogeneity of storage technologies potentially available to scientific applications, such as burst buffers, managed cloud parallel file systems (PFS), and object stores. However, those applications cannot easily utilize those technologies, because they are designed for traditional HPC systems that offer very high remote storage and network bandwidth. We present iFast, a new distributed host-side logging approach to transparently accelerating scientific applications. iFast has a strong emphasis on deployability, supporting unmodified MPI applications with unmodified MPI implementations while preserving the crash consistency semantics. We evaluate iFast on traditional HPC, cloud HPC, local cluster, and a hybrid of both, using three scientific applications. iFast reduces end-to-end execution time by 13-26% for popular scientific applications on the cloud. We show for the first time, how an application on a recent production HPC system can write data to S3 storage through fully fledged MPI-IO, in a readily shareable format.
△ Less
Submitted 2 August, 2024; v1 submitted 25 January, 2024;
originally announced January 2024.
