A Distributed Device Paradigm for Commodity ... - Semantic Scholar
Recommend Documents
module, relying on tracking of visual fiducial markers. The novelty of our con- cept is the integration of computer vision, speech recognition, AR visualization,.
provides standard services such as naming, .... placement of the replicas and the mechanisms by ..... search engines keep a local cache copy of many popular ...
Jul 12, 2014 - of applications, from server-based to rich internet applica- tions. .... the different possibilities of programming using JS, but at the same ... language is valid, but also which option offers the best per- ..... desktop grid systems.
Section 7 talks about the various failure and recovery scenarios. Section 8 discusses ... Put another way, the disk-data and disk-locks can be thought of as a.
Computerised Characters, Emotions. ABSTRACT ... mindless artificial life forms and what computer ..... asked to contribute between one and one and a half.
distributed systems and then show how views can be enriched to convey structural ... process control, and collaborative work are just some of the ... For certain application classes, including mobile comput- .... Note that view synchrony does not pla
Matteo Bonifacio, Paolo Bouquet, and Alberto ... matteo . bonifacio@it . arthurandersen . com. P. Bouquet ..... practitioners.First of all, we need to take into ac-.
Device-to-Device (D2D) communications are planned to become an indispensable .... add-on layer to standard cellular communication networks, in the sense that they are .... Generic QoE provisioning framework. ⢠QoE data collection: The ...
Jan 21, 2000 - limited by the availability of sufficient, affordable computing capacity. ...... the cheap: 94 Gflops computer system at CDN$3680/gigaflop, Poster.
Just Add Wheels: Leveraging Commodity Laptop. Hardware for Robotics and AI Education. Jonathan Kelly, Jonathan Binney, Arvind Pereira, Omair Khan and ...
Jan 21, 2000 - networked PC cluster, or Beowulf, dedicated to computational chemistry ... dedicated supercomputer performance within the budget of most .... constructed and used as dedicated compute servers for computational chemistry ...
we will be able to compose complex applications for ... contained software business components in the form of ... application domain in industrial enterprises.
agents are dedicated to the evaluation of the sensors information (target .... A Java server has been hosted on the PPC for allowing the communication between ...
Jan 26, 2001 - ing transformational support between Ninja services and terminal devices. Active proxies perform data distillation, protocol adaptation, caching ...
scale applications that access a variety of high-end resources shared with others. .... middleware toolkit for the development of Grid Computing Environments.
Email: [email protected] ... of the routing problem in multi-hop wireless networks that takes into account ..... In each couple, one node acts like sender,.
2 Institute of Computer Graphics and Algorithms, Vienna University of. Technology, Austria ... so that the technical constraints are satisfied as in the OPT variant.
The Eurographics workshop series on Object-Oriented Graphics has reflected a ... ulate lists. Multidimensional arrays have to be simulated, and arithmetic is inefficient .... character combinations that are replaced by event-specific information befo
Generate code from the model, for example, code for a specific simulator. ... language Python [21], arbitrary Python code can also be used. A similar approach to ...
with Hermann von Helmholtz [23], that the central and the most pressing issue ..... justified in calling function f the quality of the (learning) class perception ([11]). ..... acd â θ abac â θ . We obtain six expanded new sets of operations (e
Feb 14, 2017 - Van Rensburg, S.; Cook-Mozaffari, P.; van Schalkwyk, D.; van der Watt, J.; Vincent, T.; Purchase, I. Hepatocellular carcinoma and dietary ...
Annamma Abraham, PhD. Professor and Head, Department of ..... Intelligent Control: An Overview of. Techniques, Chapter in: T. Samad, Ed., âPerspectives in.
In multi-formalism modelling, each system component and/or aspect represented ..... As such, a formalism describes the set of all legal models in the formalism.
A Distributed Device Paradigm for Commodity ... - Semantic Scholar
A Distributed Device Paradigm for Commodity Applica8ons
Enrique Cauich, John Duselis, Richert Wang and Isaac Scherson Dept. of Computer Science – Systems Donald Bren School of InformaAon and Computer Sciences University of California, Irvine Irvine, CA 92697
1
Observa8on (Mo8va8on) In most modern (corporate) networked system installa8ons, a large number of computers are underu8lized. Powerful machines are devoted to liFle work and/or a few non‐demanding tasks, while other processing computers (used for data analysis) may require more resources than normally implemented. 2
Conjectures Remote high performance components/devices can be aggregated and presented to a system as local virtual devices. These virtual devices can deliver be:er performance than local physical devices by using the network links as a virtual system bus. Applica>ons using virtual devices do not require any code modifica>ons. 3
Two Experimental Case Studies • High Performance Distributed RamDisk Improves I/O intensive (Disk) applica8ons’ performance by using idle semiconductor memory resources in the network.
• Distributed Graphics Device A Codec device that enhances the system with ray tracing capabili8es by adding virtualized idle processor and graphics devices.
9
Distributed RamDisk • Harnessing available resources from under‐u8lized systems in the network. • Memory is an important resource. • Generally used as temporary storage to speed up memory references
Driver Interface Consolidator Resource Discovery
Improves I/O intensive (Disk) applica8ons performance by using idle memory resources in the network.
High Speed Network VComponent
Storage bandwidth is increased and storage latency is decreased
Resource Manager
RAM
VComponent
RAM
VComponent
RAM
VComponent
RAM
10
Distributed RamDisk Read Throughput in MB/s 1 KB
2 KB
4 KB
8 KB
16 KB
32 KB
64 KB
100 90
Throughput in MB/s
80 70 60 50 40 30 20 10 0 0
4
8
12
16
20
24
28
32
Outstanding Requests 11
Distributed RamDisk Write Throughput in MB/s 1 KB
2 KB
4 KB
8 KB
16 KB
32 KB
64 KB
90
Throughput in MB/s
80 70 60 50 40 30 20 10 0 0
4
8
12
16
Outstanding Requests
20
24
28
32
12
Distributed RamDisk Comparison using 8 outstanding requests 100 90
Throughput in MB/s
80
10k rpm read
70
10k rpm write
60
7.2k rpm read
50
7.2k rpm write
40
Hinnes
DRD 1 read
30
DRD 1 write
20
Roussev Roussev
10
Hinnes
0 1
2
4
8
16
32
64
128
256
512
Request Block Size
13
Accordion Effect TCP S
egm
TCP Send Buffer
ents
X X
ACKs s t n e gm
e TCP S d e t a Repe
TCP S
TCP Receive Buffer
egm
TCP Send Buffer
ent R etran
smis
TCP
d smiFe n a r t e R
ACK s t n e Segm
sion
TCP Receive Buffer
TCP Send Buffer Next TCP Segme
nt
Resource Lender
Resource Borrower
14
TCP receive window size
15
TCP receive window size
16
TCP receive window size (Cont.) Distributed RamDisk throughput while using 1, 2 and 3 of lenders 100
80 70 60 50 40 30 20 10
8
1 Memory Lender 64 KB TCP RcvWnd
16
2 Memory Lenders 32 KB TCP RcvWnd
64 KB
32 KB
16 KB
8 KB
4 KB
2 KB
1 KB
64 KB
32 KB
16 KB
8 KB
4 KB
2 KB
1 KB
64 KB
32 KB
16 KB
8 KB
4 KB
2 KB
0
1 KB
Throughput in MB/s
90
32
3 Memory Lenders 24 KB TCP RcvWnd
17
Database Applica8ons’ Performance Improvement Given: • a distributed RamDisk 1 borrower/3 lenders configura8on interconnected by a Gigabit switch • SQLServer 2008 • TPC‐H benchmark
Index File Access
Conjecture: a database applica8on can improve its performance by storing index files in a distributed RamDisk. The database does not require any code modifica8on.
18
Database Performance Improvement Time spent by the database performing the index search Request Size:
Second Case Study: A Distributed Codec Device Win32 Applica8on
Windows Image Component
Codec Device
High Speed Network
Graphics Hardware/So^ware
Virtual HAL VComponent
CPU
VComponent
Graphics Card
22
Second Case Study: A Distributed Codec Device • Consolidator par88ons the request. • Resource manager calculates the suitability of each virtual component
Driver Interface Consolidator Resource Discovery
Resource Manager
High Speed Network VComponent
CPU
VComponent
VComponent
Graphics Card
Graphics Card
VComponent
CPU
23
Codec Results Image decodificaAon Ame using different number of lenders and different load distribuAon mechanisms 700000
Conclusions • A new level of abstrac8on, namely the Device Level, and a new level of implementa8on, namely the Driver Level, are added to the virtualiza8on taxonomy. • The Device Abstrac8on Level increases system performance capacity by aggrega8ng idle networked resources and binding them as a single device. • The Driver Implementa8on Level provides support for legacy applica8ons by avoiding code modifica8ons. 25
