课题研究报告

Size: px

Start display at page:

Download "课题研究报告"

累冠
5 years ago
Views:

1 An Adaptive Distributed Load Scheduling Schema Bin Cheng Hai Jin Ke Shi (Compute Science and TechnologySchool of HUST, Wuhan, ) AbstractA distributed Load balancing system is designed in this paper, which overcomes the single point of failure and the performance bottleneck existing in the central scheduling system. The system uses a novel adaptive distributed schedule mechanism to complete the dynamic load balancing for the requests from the clients, which is not only dynamic adjusted according the various load, but also needs not the communication among the nodes. So the system has a good self-adapted function. In addition, the system is implemented in the Linux kernel level, which is transparent to the client and server applications and whose management and configurations are very convenience and easy. So the system is novel and practical. Key words: competitive delay, adaptive scheduling, distribute scheduling, load balancing 1. Internet Web CPU CGICommon Gateway InterfaceWeb [3][5] EDDIE Reverse-Proxy pweb WEB 1

2 imac C A x 1x 8x 2x 9x 3x A 10x 4x 11x 5x 12x 6x 7x 1x 8x 2x 9x 3x B 10x 4x 11x 5x 12x 6x HTTP WebIBM TCP RouterNetwork DispatcherLinux Virtual Server [4] LinuxIP Cisco LocalDirector TCP IP TCP [2][6] [9] 2. 1 HUB IP VIP IP IP Ethernet 1 Fig. 1 Architecture 2

3 3. 2 Server TCP Server Server 1 t1 VIPSYN [1] 2 SYN HUB Server 3 ServerSYN ServerServer ServerSYN+ACK Load A < Load B Delay A < Delay B A 4 t2 Load Server A VIP 5 t3 Load Server B t3 > t2 TCP 6 HUBServerACK N ACK N ISN + 1 TCPRSTServer A Server BServer A Load A Delay A Load B Delay B 2 Fig. 2 The Flowchart of Scheduling 3

4 [7] IPIPMAC MACIP ARP IP MAC IP ARP MAC ARP MAC IP MAC ARP MAC IP ARP IP MAC IP ARP /net/ipv4/arp.c ARP MAC IP UDP ICMP ICMP /net/ipv4/udp.c udp_recv() ICMP ifconfig MAC MAC 00:03:61:83:2a:8b 1 eth1 [root@node1 /]# ifconfig eth1 down [root@node1 /]# ifconfig eth1 hw ether 44:63:61:83:2f:8e [root@node1 /]# ifconfig eth1 up 4.2. TCP 3 SYN CPU 3 Fig. 3 The Module for Delaying 4 TCP

5 1) 2) CPU LOAD CPU, LOAD Disk, LOAD Mem, LOAD NIC 3) Delay (i) f (i) ( Load (i) ) RTT MAX ( Load (i) / Load MAX ) Load (i) Load CPU(i) Weight CPU Load Disk(i) Weight Disk Load Mem(i) Weight Mem Load NIC(i) Weight NIC Load MAX 1 Load CPU(i) ( CPU used(i) / CPU total(i) ) 100 Load Disk(i) ( IOBW used(i) / IOBW total(i) ) 100 Load Mem(i) ( MEM used(i) / MEM total(i) ) 100 Load NIC(i) ( NICBW used(i) / NICBW total(i) ) 100 Weight CPU Weight Disk Weight Mem Weight NIC Table 1 Illumination of Symbols i Delay (i) RTT MAX, i = A, B, C, D i RTT MAX Hop Round-Trip Time Load MAX 1 Load (i) CPU used(i) i i CPU CPU total(i) CPU total(i) = CPU idle(i) + CPU used(i) = 1 IOBW used(i) IOBW total(i) MEM used(i) MEM total(i) NICBW used(i) NICBW total(i) DISK IO DISK IO i i i i Weight CPU Weight Disk Weight Mem Weight NIC Weight CPU Weight Mem Weight NIC 0Weight Disk 1 Weight CPU 0.5Weight Disk 0.5Weight Mem Weight NIC 0 Weight CPU Weight Mem Weight NIC Weight Disk ) SYN SYN TCP 5) SYN 5

6 PASS TCP 4 Fig. 4 The Module for Packet Filtering 1) PASS IP IP TCP/IP 2) PASS PASS SYN SynAck ACK RST 3) PASS 4) PASS TCP TCP IP IP PASS PASS 4.4. TCP TCP socket DoS proc syn-cookie 6

7 Data General Data General Data General Data General 3Com M Web Web httperf httpd CPU MonitorAgent CPUGather PC Admin 4 A,B,C,D 100M 100M 2 Table2 The Configuration of Testing Environment CPU RAM NIC DISK (HZ) M Mbps A P4 1.4G MAXTOR Linux B PIII 550M MAXTOR Linux C Celeron 333M MAXTOR Linux D P4 1.7G MAXTOR Linux PIII 550M MAXTOR Linux COM 3C /100M HUB 3COM 100Mbps HUB 2 5 Server VIP IP ( ~ ) 5 Fig. 5 The Architecture of Testing System 7

8 3 PASS PASS 3 SYNSynAckRST TCPUDPICMPANY ANY SYNSynAckRSTANY IP / IP / IP / IP / PASS DROP 3 WEB IP Table3 An Example of Configuration IP IP TCP SYN ANY ANY PASS TCP SynAck ANY ANY PASS TCP RST ANY ANY PASS ANY ANY ANY ANY ANY PASS ANY ANY ANY ANY ANY ANY DROP 4 CPU 4 4 Table 4 The Weight Allocation for Different Servers Weight CPU Weight Mem Weight NIC Weight Disk HTTP FTP RSTP ABCD Admin CPUGather Monitor Server Agent httperf 700 Monitor Start Agent Buffer Monitor Get Agent Buffer Monitor 6 30 CPU 8

9 CPU Utilization (%) A B C D Mean Time (second) 6 CUP Fig. 6 The CPU Utilization of All Nodes ABCD Server CPU Server LVSRRONE-IP [8] ServerABCDA LVSLVSDRONE-IPBroadcast 5 5,LVSONE-IPRR ABCDCPUADLB CPU 5 LVSONE-IPADLB CPU Table 5 The Comparison of CPU Utilization Arrive Rate Reply Rate Type Mean CPU Utilization (%) (Conn/s) (Conn/s) A B C D LVS CPU ONE-IP CPU ADLB CPU AServer 78I/O Web2976Bytes 71Node 911 conn/s2node1727 conn/s3node2504 conn/s, 4Node3183 conn/s914 9

10 8I/O1Node 23.4Mbps2Node43.8Mbps3Node63.4Mbps481.4Mbps 10I/O14 I/O Reply Rate (conns/sec) NODE 2NODES 3NODES 4NODES Aggregate Net I/O BandWidth (Mbps) NODE 2NODES 3NODES 4NODES Connection Rate (conns/sec) Arrive Rate (conns/sec) 7 Fig. 7 Throughput 8 I/O Fig. 8 Aggregate Net I/O Reply Rate (conns/sec) Max Aggregate Net IO BW (Mbps) Nodes Number Nodes Number 9 Fig. 9 Max Throughput 10 I/O Fig. 10 Max Aggregate Net I/O Byte IP MTU 1500 FTP RTSP 1 10

11 6 Table 6 Overloaded Response Time Response Timems Min Max Mean A Server A Server A Server SYN [] [1] Richard W. StevensTCP/IPIII2000 [2] Web (6) [3] NOW ALBSIN [4] Internet (1) [5] (1)25~29 [6] (10) [7] Sujit Vaidya and Kenneth J. Christensen, A Single System Image Server Cluster using Duplicated MAC and IP Addresses, Proceedings. LCN th Annual IEEE Conference on, Nov. 2001, Pages: [8] Om P. Damani, P. Emerald Chung, Yennun Huang. ONE-IP: Techniques for Hosting a Service on a Cluster of Machines. the sixth international conference on World Wide Web, 1997 [9] Valeria Cardellini, Michele Colajanni. Dynamic load balancing on Web-Server Systems. IEEE Internet Computing,

12 Reference [1] Richard W. StevensTCP/IP illuminationbeijing: machine & industry press2000 [2] Fan GC, Zhu H, Huang T, Feng YL. Towards Adaptive Load Balancing Services for Web Application Servers. Journal of Software, (6):1134~1141. [3] CHEN Ximing, LU Xianliang. Research and implementation of a load balancing system named ALBSIN in NOW, Journal of Mini-Micro Systems, 2001,22(9): [4] ZHANG Wen-song, WU Ting-ting, JIN Shi-yao, WU Quan-yuan. Design and Implementation of a Virtual Internet Server. Journal of Software, 2000, 11(1): [5] CHEN Hua-ping, J I Yong-chang, CHEN Guo-liang. A Universal Model of Distributed Dynamic Load Balancing. Journal of software. 1998,9(1):25~29. [6] Lin Cheng-jiang, Li san-li. Strategy and Simulation of ARID for Distributed Dynamic Load Balancing in parallel systems. Journal of computer, 1995, 18(10): [7] Sujit Vaidya and Kenneth J. Christensen. A Single System Image Server Cluster using Duplicated MAC and IP Addresses, Proceedings. LCN th Annual IEEE Conference on, Nov. 2001, Pages: [8] Om P. Damani, P. Emerald Chung, Yennun Huang. ONE-IP: Techniques for Hosting a Service on a Cluster of Machines. the sixth international conference on World Wide Web, [9] Valeria Cardellini, Michele Colajanni. Dynamic load balancing on Web-Server Systems, IEEE Internet Computing,

ch09.PDF

ch09.PDF 9-1 / (TCP/IP) TCP/IP TCP/IP ( ) ICMP ARP RARP 9.1 TCP/IP 9.1.1 TCP/IP OSI TCP/IP (DARPA) DARPA TCP/IP UNIX Berkeley Software DistributionTCP/IP TCP/IP TCP/IP TCP/IP TCP/IP TCP/IP OSI - TCP/IP ( ) ( )