Experiment for Multi-client and Multi-server _MCMS_ System for P2P by hcj

VIEWS: 5 PAGES: 6

									Experiment of a Multi-client and Multi-server (MCMS) System for P2P Computing
Jun Ni, Ph.D.M.E. AT-RS, Information Technology Services, The University of Iowa Iowa City, IA 52242 jun-ni@uiowa.edu

Experime nt ID: MCMS03_integration (Newton-Cote Numerical Integration) Experime nt objective: Pass multiple data of double type to servers (computational nodes) for intensive computation. Return the computational result back to client (service node) for accumulation. It is a very simple linear decomposition example. Study the performance by calculating and plotting the speedup vs. the number of servers, and degree of intensive computation of server nodes. This version is different from the previous one is it is not GUI based and code is improved in much more efficiency. The input variables are a, b, and total global n, while local n is determed. Experime nt strategy: Apply Newton-Cote numerical algorithm for a function integration described as

I   f ( x)dx   wi f ( xi )  Rn
a i 0

b

i n

The a and b are the low and upper limits. Rn stands for remainder. f(x) is a function, wi (i=1,2,3,…n) is a weight coefficient, which depends repeated number and equalspaced discretization of sample points (xi). If sample points can be selected as 1, 2, or 3, which respectively represent rectangular, trapezoid, and Simplson’s methods for numerical integration. One can decomposition whole region [a,b] into multiple sub-regions, and each subregion’s integration is subject to compute on each server node. After server node calculate the sub-integration and return back to client for summation. That is

bj

Ij 



f ( x)dx   wi f ( xi )  Rn
i 0

i n

aj

xi 

bj  a j nj

(j=1,2,3,…N)

, x j  a j  xi

a, b and nglobal are given. aj, bj, and nlocal, are the local (sub-regional) low and upper limits, and repeated number for each sub area during server’s numerical computation. N is the total number of sub-regions, which can be selected as the number of servers deployed in the test. Therefore, the total number of repeated number is nj =nglobal/N , which should be used for sequential computation.
ba j N ba bj  a j  N aj  a 

For each sub- integral, the client transfers three values (aj, bj, and nj) through the Internet to servers. After each server calculates each sub-integral, it returns the value for client. The client accumulates the sub- value and come up the total value of integration. The result is also compared with sequential computing performed on client along.
I  Ij
j 1 N

The results obtained from the parallel computation on the multiple servers and sequential computation on the client along are compared for understand the performance of MCMS system. Experime nt Designer: Jun Ni Experime nt Date: July 29, 2003 Experime nt Equipments: Total 14 Macs with G3 processors Dynamic IPs as Server nodes (13 computation node) 218.255.163.92 HD) (MacOSX, G3, 350 MHz/1 MB cache, 128 MB, SDRAM 12 GB

218.255.163.135 HD) 218.255.163.106 HD) 218.255.163.219 HD) 218.255.163.227 HD) 218.255.163.228 HD) 218.255.163.237 HD) 218.255.163.16 HD) 128.255.163.242 HD) 128.255.163.96 HD) 128.255.163.133 HD) 128.255.163.136 HD) 128.255.163.141 HD)

(MacOSX, G3, 350 MHz/1 MB cache, 64 MB, SDRAM 6 GB (MacOSX, G3, 350 MHz/1 MB cache, 128 MB, SDRAM 12 GB (MacOSX, G3, 350 MHz/1 MB cache, 128 MB, SDRAM 12 GB (MacOSX, G3, 350 MHz/1 MB cache, 128 MB, SDRAM 12 GB (MacOSX, G3, 350 MHz/1 MB cache, 256 MB, SDRAM 6 GB (MacOSX, G3, 350 MHz/1 MB cache, 128 MB, SDRAM 12 GB (MacOSX, G3, 350 MHz/1 MB cache, 128 MB, SDRAM 12 GB (MacOSX, G3, 350 MHz/1 MB cache, 128 MB, SDRAM 12 GB (MacOSX, G3, 350 MHz/1 MB cache, 256 MB, SDRAM 12 GB (MacOSX, G3, 350 MHz/1 MB cache, 128 MB, SDRAM 12 GB (MacOSX, G3, 350 MHz/1 MB cache, 256 MB, SDRAM 12 GB (MacOSX, G3, 350 MHz/1 MB cache, 128 MB, SDRAM 12 GB

Dynamic IPs as 1 client node (job distribution node, or service node, which performs decomposition and job distribution) 218.255.163.101 (MacOSX, G3, 350MHz/1 MB chache, 64MB, SDRAM 6GB HD) 100 Bits Per Second Switch (3Com Super stack 3300x24 ports)

Experime ntal architecture: MCMS architecture

Job distribution and scheduling node (Client architecture)

Computational nodes (server architecture)

Each node contains a serverSocket accepts socket in a multithread Multiple sockets of each connects an individual computational node

Experime ntal Programs:
Contact Dr. Jun Ni

Experime nt Results: See attached MS Excel file There are two major studies (during the day-time and night-time) Each case, we study the intensity of decomposition. During day time:

10 9 8 7 n=1,000 n=10,000 n=100,000 n=1,000,000 n=10,000,000

speedup

6 5 4 3 2 1 0 0 2 4 6 8 10 num ber of com putational nodes (servers)

During the night time:
10.0 9.0 8.0 7.0

speedup

6.0 5.0 4.0 3.0 2.0 1.0 0.0 0.0 5.0 10.0 number of computational nodes (servers)

n=100,000,000 n=10,000,000 n=1,000,000 n=100,000

Conclusion:

1. MCMS system can be considered as a feasible distributed system for scientific computing. It can be applied to real applications. 2. More servers, the better the performance, especially for intensive computation on each server node. 3. There should be several important variables such as a. ratio of individual communication time between each client and server vs. server computation time for each task b. individual communication data transfer rate between client and server (dependent upon the simultaneous bandwidth on network) c. ratio of global communication time vs. global computation time. 4. Large physical domain can be decomposed into many small sub-domains. Each sub task for sub-domain is subject for server’s computation. 5. The speedup is dependant upon the computational intensity. The large computation on each server the better result of performance. 6. Network impacts the total computation time, since we use TCP/IP through DNS 7. When computation is less intensive or more communication between servers and client, the benefit is less 8. Each computational node may have different time of computing. Each global job’s computing varies, enhanced the total job time is varied. The result is frustrating due to networking effects and loading of server nodes. 9. Network computing is important issue. Delay due to networking is also very important issue. 10. Different time may have different result in terms of total job accomplished time 11. Statistical analysis of multi variables is need to evaluation the networking portion to the total computing. 12. The reliability of networking based computing is a major subject to study. 13. Job scheduling is temporary interval. The computational environment varies all the time. Suggestions of next experime nt: Local networked rather than DNS networked. Statistical analysis of each performance. Establish some fundamentals for analyzing grid enhanced computing.


								
To top