Upload
ajay-patil
View
215
Download
0
Embed Size (px)
Citation preview
7/30/2019 Dos Presentation
1/28
Application of Self OrganizingMaps for Investigating Network
Latency on Broadcast-basedDistributed Shared Memory
Multiprocessor
7/30/2019 Dos Presentation
2/28
INTRODUCTION
Distributed shared memory(DSM)
multiprocessor, which represent a successful
hybrid of two parallel computer classes :-1. Shared memory multiprocessor
2. Distributed computer systems.
7/30/2019 Dos Presentation
3/28
DSM multiprocessor systems are an
attractive and viable platform for both Parallel
and Mainstream computing.Due to its advantages in terms of
1. Scalability
2. Programmability
3. Cost effectiveness
7/30/2019 Dos Presentation
4/28
Aim of DSM system
To combine of PROGRAMMING of shared
memory multiprocessors with the
SCALABILITY of distributed computersystems.
7/30/2019 Dos Presentation
5/28
DSM system
7/30/2019 Dos Presentation
6/28
Remote Memory Latency
To achieve high performance in DSM
multiprocessor Remote Memory Access time must
be minimized
RML is caused by accessing a memory location in a
processor other than the one originating the request
Network latency is major component of remote
memory latency
Cause most of the time is consumed in
communication over the interconnection network
7/30/2019 Dos Presentation
7/28
Developing methods to minimize the network
latency in broadcast-based optical DSM
multiprocessors requires the knowledge ofthe effect of DSM parameters on the network
latency
7/30/2019 Dos Presentation
8/28
Self-Organizing
Maps
7/30/2019 Dos Presentation
9/28
Overview
Self organizing map (SOM) is a data visualization techniqueinvented by kohonen
SOM is a way to represent higher dimensional data in anusually 2-D or 3-D manner, such that similar data is groupedtogether.
It runs unsupervised and performs the grouping on its own.
SOM is accepted as Analysis tool.
7/30/2019 Dos Presentation
10/28
SOM Structure (kohonen)
Every node is connected to the input the same way, and nonodes are connected to each other.
In a SOM that judges RGB color values and tries to group similar
colors together, each square pixel in the display is a node inthe SOM.
Notice in the converged SOM above that dark blue is near lightblue and dark green is near light green.
7/30/2019 Dos Presentation
11/28
We use SOM to investigate the realtionship
between the network latency and several
DSM parameters on the SimultaneousOptical Multiprocessor Exchange
Bus(SOME-Bus)
SOME-Bus is high bandwidth, optical and
broadcast-based interconnection network
supporting DSM.
7/30/2019 Dos Presentation
12/28
An event based discrete simulator
Is developed for simulating a SOME-Bus
system Which containing 64 nodes,
Each of which has Processor, a Directorycontroller, A cache controller, and an output
channel.
7/30/2019 Dos Presentation
13/28
Node model of DSM-basedSOME-bus system
7/30/2019 Dos Presentation
14/28
State diagram of acb_fifo inOPNET modeler
initarrival
startcomp
l
idle
arrival
arrival
default
default
default
Queue empty
Svc completion
Server busy & insert ok
7/30/2019 Dos Presentation
15/28
Dataset details
Two thousand data points have been collected in order
to create a dataset for study.
Dataset includes 6 parameters
1. Miss rate to a modified block (Pm)2. Fraction of write misses (Pw)
3. Probability of having an upgrade ownership request message
(Puor)
4. Probability of having a cache full (Pcf)
5. Ratio of the mean thread run time to mean message transfer time
(T/R)
6. Network latency
7/30/2019 Dos Presentation
16/28
In observed simulation
Pw ranges from 0.15 to 0.7
Pm takes values varying between 0.2 and0.6
The ration varies between 0.05 and 1
7/30/2019 Dos Presentation
17/28
Let m =miss rate F= number of instruction
per second performed by the processor at
each node, S= mean msg size in bytes andC= channel bandwidth(bytes/s)
Ratio of the mean thread run time to the
mean message transfer time T/R=mSF/C
7/30/2019 Dos Presentation
18/28
7/30/2019 Dos Presentation
19/28
7/30/2019 Dos Presentation
20/28
7/30/2019 Dos Presentation
21/28
7/30/2019 Dos Presentation
22/28
7/30/2019 Dos Presentation
23/28
7/30/2019 Dos Presentation
24/28
7/30/2019 Dos Presentation
25/28
7/30/2019 Dos Presentation
26/28
7/30/2019 Dos Presentation
27/28
CONCLUSION
This paper presents a new application domain
for the use of SOM.
In this paper SOM have been used to
investigate the ralation between network latency
and important DSM parameters
The result shows that network latency has
highest correlation with T/R while Pcf has theleast effect on the values of network latency
7/30/2019 Dos Presentation
28/28
Puor requires high Pw values to
demonstrate strength on network latency
values. Low values of Pw and Pm will yield low
network latencies
Low values of Pw together with high valuesof Pm have a detractive effect on network
latency