p-Jigsaw: A Cluster-based Web Server with Cooperative Caching Supports: p-Jigsaw: A Cluster-based Web Server with Cooperative Caching Supports Ge Chen, Cho-Li Wang, Francis C.M. Lau (Presented by Cho-Li Wang) The Systems Research Group Department of Computer Science and Information Systems The University of Hong Kong


What’s a cluster ?: What’s a cluster ? A cluster is a type of parallel or distributed processing system, which consists of a collection of interconnected stand-alone/complete computers cooperatively working together as a single, integrated computing resource – IEEE TFCC.


Rich Man’s Cluster: Rich Man’s Cluster Computational Plant (C-Plant cluster) Rank: 30 at TOP500 (11/2001) 1536 Compaq DS10L 1U servers (466 MHz Alpha 21264 (EV6) microprocessor, 256 MB ECC SDRAM) Each node contains a 64-bit, 33 MHz Myrinet network interface card (1.28 Gbps/s) connected to a 64-port Mesh64 switch. 48 cabinets, each of which contains 32 nodes (48x32=1536)


Poor Man’s Cluster: Poor Man’s Cluster HKU Linux Cluster 32 733 MHz Pentium III PCs, 392MB Memory Hierarchical Ethernet-based network : four 24-port Fast Ethernet switches + one 8-port Gigabit Ethernet backbone switch) Additional 80-port Cisco Catalyst 2980G Fast Ethernet Switch


Cluster vs. Supercomputer: Cluster vs. Supercomputer Supercomputer: a HPC system capable of efficiently processing large-scale technical computing problems. E.g., ASCI machines, IBM Blue Gene. Cluster: a HPC system that integrates mainstream commodity components to process large-scale general problems, including technical computing, business applications, and networking services.


Cluster Computer Architecture: Cluster Computer Architecture High-Speed LAN (Fast/Gigabit Ethernet, SCI, Myrinet) Availability Infrastructure Single System Image Infrastructure Programming Environment (Java, C, MPI, HPF) Web Windows User Interface Other Subsystems (Database, Web server OLTP, etc.) OS Node OS Node OS Node OS Node


Talk Outline: Talk Outline Motivation -- The Need for Speed Cluster-based Solutions System Architecture of p-Jigsaw Performance Evaluation Conclusion and Future Work Other SRG Projects


The Challenges: The Challenges Netscape Web site in November 1996: 120 million hits per day Microsoft Corp. Web site received more than 100 M hits per day. (1,200 hits per second) Olympic Winter Games 1998 (Japan): 634.7M (16 days), peak day 57M, peak minute 110K Winbledon July 1999, 942 M hits (14 days), peak day 125M, peak minute 430K Olympic Games 2000 : peak day 502.6 M, peak minute 600K hits. (10K hits per second)


The Need for Speed: The Need for Speed Internet user popularity is growing very fast According to United States Internet Council’s report, regular Internet user has increased from less then 9M in 1993 to more than 300M in the summer of 2000, and is still growing fast Broadband becomes popular According IDG’s report, 57% of the workers in U.S access Internet via broadband in office. The figure will be more than 90% by 2005. Home broadband user will also increase from less than 9M now to over 55M by 2005 HTTP requests account for larger portion of Internet traffic now One study shows that HTTP activity has grown to account for 75%~80% of all Internet traffic


Internet Still Growing Dramatically: Internet Still Growing Dramatically Source: Dr. Lawrence Roberts, Caspian Networks; August 15th, 2001 Internet traffic growth rate increased from 2.8 to 4 per year in 2000. Faster than the average of 2.8 times per year since the Internet began aggressively in 1997 Internet traffic growth rate continued to grow at 4 per year through Q1 2001 Service providers have capital problems They can only avoid equipment purchases for a short period They must buy equipment soon to hold market share


The Need for Speed: The Need for Speed The Need for Speed Growing user number Faster last-mile connection speed Increasing portion of HTTP requests accounts for all Internet traffic Require a more powerful Web server architecture


Cluster-Based Solution: Cluster-Based Solution Cluster -- A Low-cost yet efficient parallel computing architecture


Cluster-based Solutions: Cluster-based Solutions Cluster-based Web Server System DNS-based Dispatcher-based (More references: V. Cardellini, M. Colajanni, and P.S. Yu, "Dynamic Load Balancing on Web-Server Systems", IEEE Internet Computing, May/June 1999, pp.28-39.)


DNS-based Approach: DNS-based Approach 2 Cluster DNS client Intermediate name servers 1 Step 1: Address request (URL) Step 1’: Address request reaches the DNS Step 2: (Web-server IP address, TTL) selection Step 3: Address mapping (URL --> IP address 1) Step 3’: Address mapping (URL --> IP address 1) Step 4: Document request (IP address 1) Step 5: Document response (IP address 1) 1’ 3’ 3 Server 1 (IP address 1) Server N (IP address N) 4 5 5


Dispatcher-based: Dispatcher-based Dispatcher A network component of the Web-server system acts as a dispatcher which routes the requests to one of the Web servers to fulfill the load balancing. Each Web server works individually. client Layer 4 switching with level 2 address translation : One-IP, IBM eNetwork, WebMux, LVS in DR mode Layer 4 switching with level 3 address translation : Cisco LocalDirector, Alteon ACEDirector, F5 Big/IP, LVS in NAT mode. Layer 7 switching (Content-based): LARD, IBM Web Accelerator, Zeus Load Balancer (ZLB)


Previous Researches Focus: Previous Researches Focus Previous researches mainly focus on: Load Distribution/Balancing [H. Bryhni et al. 2000] Scalability [Trevor Schroeder et al. 2000] High Availability [Guillaume Pierre et al. 2000] Caching policies on a single web server [Martin Arlitt et al. 2000]


p-Jigsaw -- Goals: p-Jigsaw -- Goals High Efficiency: Explore aggregate power of cluster resources (CPU, memory, disk, network bandwidth). Explore in-memory Web caching on cluster-based Web servers High Scalability Maintain high cache hit rate and high throughput as cluster size grows Eliminate potential bottleneck in the overall design High Portability Multi-platform support Heterogeneous cluster


Main Features of p-Jigsaw Web servers: Main Features of p-Jigsaw Web servers Global Object Space (GOS) Hot Objects Caching Cooperative Object Caching Distributed Cache Replacement Algorithms


Global Object Space: Global Object Space (All Web objects in system are visible and accessible to every node through GOS) Server Node Server Node Server Node Server Node High-Speed LAN Global Object Space Memory Cache Memory Cache Memory Cache Memory Cache Hot Object Cache


Hot Objects: Hot Objects Web object access pattern exhibits strong concentration Some study shows that around 10% of the distinct documents are responsible for 80-95% of all requests received by a Web server [Arlitt et al. 1997] “hot objects” refer to objects that are frequently requested in a short time interval – or “focus of interests” “Hot objects” may have duplicated copies  Improve hit rate - Avoid excessive inter-node object fetching Share the busty requests for hot objects to alleviate the workload of the home nodes holding many hot objects Realized by distributed cache replacement algorithms (To be discussed)


Construction of GOS: Construction of GOS Two tables are maintained at each node: Global Object Table (GOT) and Local Object Table (LOT) GOT keeps system wide information for mapping object URL or partial URL to their home node (HN) or servers with the cached copy (CCNN : Cache Copy Node Number), and approximated global access counter (AGAC) for cache replacement. Home node of an object refers to the node holds the persistent copy of that object


Construction of GOS: Construction of GOS LOT keeps access records for objects cached in current node’s hot object cache A copy of the approximated global access counter Local access counter (LAC) Home node of the cached object (HN)


Slide23: Incoming Request: http://p-Jigsaw.org/node4/dir2/pic1.jpg Hard disk Hot Object Cache Node 1 1 2 Node 2 Node 3 Node 4 Hard disk Hot Object Cache Hard disk Hot Object Cache Hard disk Hot Object Cache Miss! 3 4 6 HN : Home Node AGAC : Approximated Global Access Counter CCNN : Cache Copy Node Number LAC : Local Access Counter Redirect the request to node 4 (Home node of the requested page) Search on GOT1 (Hashing)


Distributed Cache Replacement: Distributed Cache Replacement Two LFU-based Algorithms are Implemented: LFU-Aging: AGAC/ 2 ; every △ t Weighted-LFU: AGAC/ (file size) Global LRU (GLRU) is implemented for comparison Try to cache the “hottest objects” in global object space Cached object’s life time is set according to HTTP timestamp. Cache consistency is maintained by invalidation scheme.


Update of Access Counters: Update of Access Counters GOS 45+200 = 245 245 245 0 LAC is periodically sent back to objects’ HN to maintain an approximate global access counter for every cached object


p-Jigsaw Implementation: p-Jigsaw Implementation A preliminary prototype system has been implemented by modifying the W3C’s Jigsaw, version 2.0.5 – all written in Java. The clients are modified version of httperf. It performs stress test on the designated Web server Test Data: Web server logs from Berkeley CS Total size: 6.756 GB Number of files : 89,689 Average file size: 80,912 bytes Number of requests: ~640,000 Data transferred: ~35GB. Distinct files requested: 52,347


Experiment Setup: Experiment Setup 32-node PC cluster. Each node consists of a 733MHz Pentium III PC running Linux 2.2.4. The nodes are connected with an 80-port Cisco Catalyst 2980G Fast Ethernet Switch. A NFS server (2-way SMP) with Gigabit Ethernet link to the switch. 16 nodes acts as clients, and the rest as Web servers. Each of the server nodes has 392MB physical memory installed 32-node PC cluster


Experiment Results: Experiment Results Effects of Scaling the Cluster Size


Experiment Results: Experiment Results Effects of Scaling the Cache Size Aggregated Cache size for 16 nodes = 1.8% (8 MB per node), 3.6%, 7.2%, and 14.4% (64 MB per node) of the size of the data set


Effects of Scaling the Cache Size: Effects of Scaling the Cache Size The largest total cache size is about 14.4% of the data set size, the cache hit rate reaches around 88% This confirms the early research observation that around 10% of the distinct objects account for 80%~95% of all the requests the server received The approximated global LFU algorithm with cooperative caching support is effective With a relatively small amount of memory in each node used for caching hot objects, we are able to obtain a high cache hit rate which increases the whole system’s performance considerably


Analysis of Requests Handle Patterns: Analysis of Requests Handle Patterns Local Cache Object (in local memory) The server that receives the request has the requested object in its local hot object cache. Peer Node Cache Object (in remote memory) The server that receives the request does not have the requested object in its local hot object cache. The object is fetched from either the home node or a or other peer nodes. Disk Object (local or remote disk) The requested object is not in the global object space, and has to be fetched from the file server. This has the longest serving time.


Analysis of Requests Handle Patterns: Analysis of Requests Handle Patterns LFU-based algorithms show high local cache hit rates. With 64 MB cache per node, the local cache hit rate is around 60% for both Weighted-LFU and LFU-Aging,


Analysis of Requests Handle Patterns: Analysis of Requests Handle Patterns ~6.7% ~35.2% ~50% ~25% With small cache size (8MB), the cooperative cache can improve the global cache hit rate and reduce the costly file server disk access, which is a common bottleneck for a website.


Analysis of Requests Handle Patterns: Analysis of Requests Handle Patterns GLRU shows a much lower local cache hit rate than the LFU-based algorithms. GLRU does achieve nearly the same global cache hit rate as the LFU-based algorithms when the cache space is large (64 MB): 52% vs. 60% in 16-node case Local cache hit rate for LRU-based algorithms drops much faster than that for LFU-based - 20% vs. 40% in 16-node case


Experiment Results: Experiment Results For the 16-node case, the local cache hit rate for GLRU drops from around 52% to around 20%, while that for Weighted-LFU with CC only drops from around 60% to around 40%.


Conclusions: Conclusions Use of cluster wide physical memory as object cache can lead to improved performance and scalability of Web server systems With relatively small amount of memory dedicated for object content caching, we are able to achieve a high hit rate with the cooperative caching Favor replicating more hot objects rather than squeezing more different objects into the global object space.


Future Work: Future Work The HKU “Hub2World” Project Build a giant proxy cache server on a large PC cluster with HKU’s 300-node Gideon cluster based on p-Jigsaw Cache hot objects on 150 GB in-memory cache (0.5GB x 300) + 12 Terabytes disk space (40GB x 300) Design of new caching algorithms


Other SRG Projects: Other SRG Projects Welcome to download our software packages and test them on your clusters. URL: http://www.srg.csis.hku.hk/


Current SRG Clusters: Current SRG Clusters


JESSICA2 – A Distributed JVM: JESSICA2 – A Distributed JVM


JUMP Software DSM: JUMP Software DSM Allows programmers to assume a globally shared virtual memory, even they execute program on nodes that do not physically share memory The DSM system will maintain the memory consistency among different machines. Data faulting, location, and movement are handled by the DSM. Proc 1 Mem 1 Proc N Mem N Proc N-1 Mem N-1 Proc 2 Mem 2 Network Globally Shared Virtual Memory


HKU DP-II on Gigabit Ethernet: HKU DP-II on Gigabit Ethernet Single-trip Latency Test (Min: 16.3 µs) Bandwidth Test (Max: 79.5 MB/s) RWCP GigaE PM : 48.3 us round-trip latency and 56.7 MB/s on Essential Gigabit Ethernet NIC Pentium II 400 MHz. RWCP GigaE PM II : 44.6 us round trip time. 98.2 MB/s bandwidth on Packet Engines G-NIC II for connecting Compaq XP-1000 (Alpha 21264 at 500 MHz.),.


SPARKLE Project: SPARKLE Project A Dynamic Software Architecture for Pervasive Computing – “Computing in Small” Won’t Fit Application Applications distributed as monolithic blocks Our component-based solution Facets


SPARKLE Project: SPARKLE Project Facet Servers Execution Servers Computational Grid Co-operative Caching (User Mobility) Intelligent proxies Facet Query Facet Retrieval Delegation/ Mobiie code Peer-to-Peer Interaction Clients (Linux + JVM) Overview of the proposed software architecture


ClusterProbe : Cluster Monitoring Tool: ClusterProbe : Cluster Monitoring Tool


Q&A: Q&A For more information, please visit http://www.csis.hku.hk/~clwang