JICOS A Java-Centric Network Computing Service

1
JICOSA Java-Centric Network Computing Service

• Peter Cappello Christopher James Coakley
• Computer Science
• University of California, Santa Barbara

2
API Goals

• Application program is oblivious to
• Number of processors
• Processor topology
• Inter-process communication
• Faulty compute servers

3
API
f(4)

• Divide Conquer (DAC)

f(3)
f(2)
f(2)
f(1)
f(1)
f(0)

f(1)
f(0)



4
API
f(4)

• DAC
• Common environment
• Input object
• Shared object

f(3)
f(2)
f(2)
f(1)
f(1)
f(0)

f(1)
f(0)



5
Architectural Goals

• Scalable
• Heterogeneous processors OS
• Mobile code
• Support adaptively parallel computation
• Tolerate faulty compute servers
• Reduce or hide communication latency

6
Architecture
H
H
H
H
H
H
S
H
H
H
H
H
H
7
Architecture
M
login setComputation getResult logout
C
8
Hiding Communication LatencyTask Caching
f(4)
f(3)
f(2)
f(2)
f(1)
f(1)
f(0)

f(1)
f(0)



9
Hiding Communication LatencyTask Pre-fetching
f(4)
f(3)
f(2)
f(2)
f(1)
f(1)
f(0)

f(1)
f(0)



10
Hiding Communication LatencyExecute Task on
Server
f(4)
f(3)
f(2)
f(2)
f(1)
f(1)
f(0)

f(1)
f(0)



11
Tolerating Faulty Hosts
Transactions kill performance
12
Tolerating Faulty Hosts
Transactions kill performance
13
Tolerating Faulty Hosts
Proxy
H
TASK
TASKS
Proxy
H
TASKS
14
Performance Experiments

• Problem 200-city TSP
• 61,295 BranchAndBound Tasks (2.05s)
• 30,647 MinSolution Tasks (lt 1ms)
• 120-processor experiments use 3 processor types
  (CX journal paper derives formula)

15
(No Transcript)
16
Fault Tolerance Experiments

• Problem 200-city TSP
• Killed p processors after 1,500s, for p 2, 6,
  20, 24, 26, 30.
• overhead
• actual time / ideal time

S
32 PROCESSORS
17
Fault Tolerance Experiments
18
Task Server Overhead
S/H
S
H
22 PROCESSORS 11 HOSTS, 1 SERVER 12
MACHINES TIME 3114.8s
22 PROCESSORS 11 HOSTS, 1 SERVER 11
MACHINES TIME 3115.1s
19
Conclusions

• API Cilk Common Task Environment
• Architecture
• network of servers, each serving many hosts.
• Supports adaptive parallelism
• Efficiently tolerates faulty hosts
• Excellent speedups
• 2 processors (1 Host) 9 hours and 32 minutes
• 120 processors lt 12 minutes (96.66 ideal)
• 3 application-controlled latency-hiding
  directives
• Small Server overhead Run Host on Server

20
THANK YOU!

• URL cs.ucsb.edu/projects/jicos
• Download
• System
• Source
• Tutorial

21
(No Transcript)
22
A Distributed Computing Taxonomy
NOT
Application fixes processor topology a priori
Adaptively parallel
NOT
Tolerates faulty computer servers
NOT
Divide Conquer API
NOT
23
Ancestry
Cilk
Linda
Atlas
Satin
Javelin
CX
Pirannha
JICOS