PRESS: a clustered server based on user-level communication

Abstract

Abstract-In this paper, we propose and evaluate a cluster-based network server called PRESS. The server relies on localityconscious request distribution and a standard for user-level communication to achieve high performance and portability. We evaluate PRESS by first isolating the performance benefits of three key features of user-level communication: low processor overhead, remote memory accesses, and zero-copy transfers. Next, we compare PRESS to servers that involve less intercluster communication, but are not as easily portable. Our results for an 8-node server cluster and five WWW traces demonstrate that user-level communication can improve performance by as much as 52 percent compared to a kernel-level protocol. Low processor overhead, remote memory writes, and zero-copy all make nontrivial contributions toward this overall gain. Our results also show that portability in PRESS causes no throughput degradation when we exploit user-level communication extensively.

Barely Alive Servers: Greener Datacenters Through Memory-Accessible, Low-Power States

“…Cluster-wide cooperative main-memory caching (or simply cooperative caching) [Pai et al(1998), Carrera and Bianchini(2005), Dormando(2011)] improves the performance of Internet service, as it caches the most popular objects in the server memories, thereby avoiding disk accesses or content regeneration. The mapping of objects to server memories is known to the intra-cluster request distribution algorithm.…”

“…However, existing cooperative caching layers differ in how they distribute the incoming client requests across the cluster. In the layer we study in this paper [Carrera and Bianchini(2005)], when a client request arrives, the distribution algorithm directs it to one of the servers that caches the requested object (if one exists), as long as that does not excessively imbalance the load across servers. When it does, the caching layer creates an additional replica of the object to better spread the load.…”

“…To be most useful, the barely-alive family requires the cluster software to have the ability to (1) consolidate the workload into a subset of (active) servers and (2) perform remote memory (read and/or write) accesses to barely-alive servers. For Internet services, it would be natural for the cluster software to implement some sort of cooperative main-memory caching system [Carrera and Bianchini(2005), Dormando(2011), Pai et al(1998)], which would manage the main memories of the cluster as a single large cache. This implementation could be coupled with a standard consolidation algorithm.…”

“…Our middleware implements the PRESS cooperative main-memory caching system [Carrera and Bianchini(2005)], but modifies it to accommodate servers in a barely-alive state and to re-size the local caches dynamically. The goal is to reach a target cache hit ratio, while allowing energy conservation and freeing up as much memory as possible for applications.…”

Distributed active element in 10 Gbps network

“…Another area related to this paper is utilization of computer clusters as either distributed routers or distributed servers [11], [12], [13]. Project Suez [14] is a distributed router based on commodity PC cluster with Myrinet interconnection with each node of the cluster having one internal interface to the Myrinet switch and optionally one or more external interfaces.…”