没有合适的资源?快使用搜索试试~ 我知道了~
首页论文《Linux for High Performance and Real-Time Computing on SMP system.pdf》
资源详情
资源评论
资源推荐
Linux for High Performance and Real-Time Computing on
SMP Systems
∗
Dominique RAGOT, Yulen SADOURNY
Thales, Colombes, France
{dominique.ragot,yulen.sadourny}@fr.thalesgroup.com
Denis FOUEILLASSAR, Philippe COUVEE
Bull, Grenoble, France
{denis.foueillassar,philippe.couvee}@bull.net
L´eonard SIBILLE
CEA List, Fontenay aux Roses, France
leonard.sibille@cea.fr
Jean-Luc DEKEYSER, Philippe MARQUET, Eric PIEL, Julien SOULA
LIFL, University of Lille, France
{jean-luc.dekeyser,philippe.marquet,eric.piel,julien.soula}@lifl.fr
Hugo KOHMANN
Dolphin Interconnect, Oslo, Norway
hugo@dolphinics.no
Alexis BERLEMONT
Openwide, Paris, France
alexis.berlemont@openwide.fr
Abstract
Applications that require a combination of high-performance computing capabilities and real-time
behavior, although pervasive (simulation, medicine, training, multimedia communications), often rely on
specific hardware and software components that make them high performance but expensive, and quite
difficult to develop, validate and moreover upgrade. The increasing performance of COTS and the volume
of software developed for these applications lead to the consideration of incremental development schemes
in addition to sole performance. In the ITEA Hyades project, industrial companies, research centres
and academic departments, propose a complete set of software technologies aimed at adding real-time
capabilities to multi-processor systems, with a strong commitment to standards. In this paper we present
the application requirements with respect to real-time, the architectural model proposed, as well as the
reasons for using the Linux operating system. Then, we introduce software components that have been
selected to provide real-time needs, among which are Adeos and ARTiS, and their expected contribution to
global performance. Finally we provide performance measurements for these elements.
∗
This work has been done in the scope of the Hyades project, ITEA 01010
1
1 Introduction
The integration of digital systems in many aspects
of life is now a reality of every day. In many fields of
activity: office, leisure, health, security, transporta-
tion, we are indeed communicating with computers,
without having to know how this communication is
managed. Terminals, computers and networks have
simply to bring together the required services to the
end-users in a seamless fashion. This integration re-
quires infrastructure components that must deliver
both functionality and performance. For a majority
of systems, performance relates to throughput, but
for a growing number of domains (video and audio
contents delivery, virtual reality, manufacturing pro-
cess control, sensor fusion) performance relates to
timely execution. Such applications have usually re-
quired non-standard and costly hardware and soft-
ware solutions. Their specificity had been for years
the justification for the use of specific technology at
all levels: specialized DSP processors, specialized
operating systems lacking the support of standard
APIs and requiring custom applications, and also
specialised cluster interconnects.
Moreover the diffusion and utilization of paral-
lel distributed systems based on COTS (components
off the shelf) technology has widely increased in last
years. Today, using COTS, it is possible to build
up powerful clusters not only for number crunch-
ing but also for highly parallel commercial applica-
tions. Many computer manufacturers have adopted
this approach, and now high performance comput-
ing systems are available at a price very low with
respect to one decade ago.
Real-time capabilities for these systems have not
reached a comparable level of maturity due to lim-
ited market size. In order to evaluate what level of
performance could be reached, a multidisciplinary
team [1] has designed and developed real-time ex-
tensions for parallel systems whose requirements,
contents, and results are exposed in the following
chapters.
2 Applications requirements
For complex applications, real-time constraints are
expressed at several levels of interaction. When
there is close man-system interactions (e.g in virtual
reality applications), the constraints are expressed in
relation to perception. On the other hand, for data
acquisition systems, the receiver/emitter must not
cause data to be lost due to lack of temporal control
over some asynchronous events.
Because they are complex, these applications
also make large usage of components that are not in
dealing at all with real-time issues. For instance all
back end processing such as classification, database
access, global configuration and monitoring, typi-
cally rely on several legacy or third-party middle-
ware and tools components. The underlying soft-
ware architecture has to provide capabilities to inte-
grate these components in a seamless fashion.
In order to assess the versatility of the proposed
architecture for this class of applications, we have
chosen the following two cases:
2.1 Virtual Reality
One application of the real-time kernel is the simu-
lation of industrial parts in virtual reality. Industri-
alists currently use real-life mock-ups for assembly
testing. This process takes a large amount of time
and money. Virtual reality makes such testing eas-
ier and cheaper. Once converted into appropriate
3D computer models, industrial parts are integrated
into a simulation framework which computes dy-
namics and collisions between parts. In addition,
the simulation is connected to a force-feedback de-
vice which enables the user to feel collision forces,
as shown on figure 1.
This device, however, must be fed with force
data at a very high rate (1kHz, typically). Failure
to respect this rate results in jitter, and eventually
makes the simulation crash. Today, the simulated
3D models can only consist of a few thousand poly-
gons, because of this rate constraint. A SMP ma-
chine enables developers to isolate and make par-
allel the dynamics and collision processes, which
should give dramatically better performance. The
real-time patch will ensure the real-time constraint
is enforced. All this should result in more detailed
models, and better testing accuracy.
FIGURE 1: Linking a haptic device to a 3D sim-
ulation
2
2.2 Video proxy
The video proxy is an application located some-
where in the network between the server and the
end-user. It is typically placed at the edge of a net-
work, where the available bandwidth or the security
requirements change (see figure 2). The purpose of
a video proxy is mainly to adapt the video streams
going through it, depending on the users’ character-
istics at the end of the delivery chain.
FIGURE 2: Proxy in video distribution
Description and Functionality The processing of
a video stream during its transmission requires spe-
cialised modules, due to the high data rates in-
volved. A video proxy allows to perform user au-
thentication as well as filtering and logging on any
traffic that traverses the proxy server. But its main
and most heavy task consists of pure video-related
processing, specifically at the edge of heterogeneous
networks:
• Transcoding of video content, i.e. dynamic
adaptation to ensure a certain quality-of-
service. Some content formats are designed to
optimise scalability, such as Motion JPEG 2000
or the upcoming MPEG-SVC, thus allowing to
transcode streams without going through the
entire encoding chain.
• Scalable encryption to ensure the confidential-
ity of critical data. This kind of encryption
selects the byte chunks to cipher and allows
to keep the structure of the video content in-
tact. One of the main advantages of these tech-
niques is to allow the transcoding of ciphered
streams.
A generic video proxy can implement some traf-
fic control, but does not contain any firewalling ca-
pability. In this way, it can be deployed behind a
traditional firewall platform. Therefore, a typical
use on a private network area can be the follow-
ing: a main firewall accepting inbound traffic, de-
termines which application is being targeted, and
then hands off the traffic to an appropriate proxy
server, e.g. videos to the video proxy. This way,
such a dedicated proxy can be used to decrease the
work load on the firewall and to perform more spe-
cialised processing that otherwise may be difficult
or even impossible to perform on the firewall itself.
Application constraints Today, the two main limi-
tations for video proxy modules are the low flexibil-
ity of content formats, although some standards are
emerging, and the computing power of networks
nodes, which have high performance for basic pro-
cesses such as routing but are not optimised for more
complex computation like transcoding.
The critical parameters for the machine when
the proxy runs are the CPU-load and the achieved
quality of service for the clients behind. The appli-
cation performs a continuous, on-stream processing
and must do the work in real-time, so that the video
quality, resolution and frame-rate remain constant
on the end-users’ players.
3 The proposed architecture
Multiprocessor systems are well suited to provide
the required processing power for such applications
as well as a choice of operating systems and mid-
dleware tools, at least when excluding real-time is-
sues. Including real-time capabilities directly usable
by application designers dramatically reduces this
choice and offers a limited set of solutions:
1. pure RTOS-based solutions are usually quite
limited in terms of middleware and tools sup-
ported, and only a very few of them have sup-
port for multiprocessor systems. The appli-
cation developer has usually no choice but to
partition the number of processors available
in two sets: one with RT capabilities running a
RTOS, and the other running a GPOS with all
applications. Communications within and be-
tween sets are done using MPI-like primitives.
Besides having to statically define resources
for RT and non-RT parts of the application,
this solution requires that all communication
software be developed in a way that is highly
dependent on the underlying machine archi-
tecture.
3
剩余11页未读,继续阅读
morre
- 粉丝: 187
- 资源: 2337
上传资源 快速赚钱
- 我的内容管理 收起
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
会员权益专享
最新资源
- 27页智慧街道信息化建设综合解决方案.pptx
- 计算机二级Ms-Office选择题汇总.doc
- 单链表的插入和删除实验报告 (2).docx
- 单链表的插入和删除实验报告.pdf
- 物联网智能终端项目设备管理方案.pdf
- 如何打造品牌的模式.doc
- 样式控制与页面布局.pdf
- 武汉理工Java实验报告(二).docx
- 2021线上新品消费趋势报告.pdf
- 第3章 Matlab中的矩阵及其运算.docx
- 基于Web的人力资源管理系统的必要性和可行性.doc
- 基于一阶倒立摆的matlab仿真实验.doc
- 速运公司物流管理模式研究教材
- 大数据与管理.pptx
- 单片机课程设计之步进电机.doc
- 大数据与数据挖掘.pptx
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功
评论2