Hardware Reference
In-Depth Information
Chapter 32
Multi/Many Core
Ramon Nou, Toni Cortes
Barcelona Supercomputing Center
Stelios Mavridis, Yannis Sfakianakis, and Angelos Bilas
Foundation for Research and Technology|Hellas
32.1
Introduction :::::::::::::::::::::::::::::::::::::::::::::::::::::: 353
32.2
Storage I/O at Present ::::::::::::::::::::::::::::::::::::::::::: 354
32.3
Storage I/O in the Near Future :::::::::::::::::::::::::::::::::: 355
32.4
Challenges and Solutions :::::::::::::::::::::::::::::::::::::::: 356
32.4.1
NUMA Eects ::::::::::::::::::::::::::::::::::::::::::: 356
32.4.2
Improving I/O Caching Eciency ::::::::::::::::::::::: 357
32.4.3
Dynamic I/O Scheduler Selection ::::::::::::::::::::::: 359
32.5
Conclusion :::::::::::::::::::::::::::::::::::::::::::::::::::::::: 360
Bibliography :::::::::::::::::::::::::::::::::::::::::::::::::::::: 360
32.1 Introduction
With current trends moving toward large multicore servers as building
blocks for HPC and data centers, storage I/O is becoming an increasing con-
cern for application scaling and performance. The main contributing factors
that lead to this landscape are: (a) technology that allows and dictates building
large multicore servers; (b) servers that are built as shared memory systems
that run a single instance of the operating system for convenience and market
size purposes; (c) with the advent of \big data" problems, applications that
tend to become more and more data intensive; and (d) fast storage devices
that allow such systems to operate at high input/output operations per second
(IOPS). Therefore, a major trend for both HPC and data centers is toward
using large shared servers as infrastructure components for computation.
Large, shared servers and data-intensive applications pose significant chal-
lenges for the storage I/O path. I/O requests need to traverse multiple layers
in the operating system, storage controllers, and the devices themselves. The
operating system in general and the Linux kernel in particular have not been
designed to handle large amounts of concurrency in the I/O path, nor to
353
