M. Aron, P. Druschel, and W. Zwaenepoel. Ecient support for P-HTTP in cluster-based web servers. In Proceedings of the USENIX 1999.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....requests for a given object across all of the servers, forcing each server to fetch the object and cache it redundantly. On the other hand, content routing policies preserve this locality by preferring the same server for repeat requests, but they are vulnerable to load imbalances. Recent research [16, 4, 5] has studied this tradeoff in depth, and developed locality aware request distribution (LARD) and related policies to combine the benefits of both approaches. We use the term URL switching to encompass URL hashing as well as the more sophisticated LARD strategies when applied in a server switch. ....

....performance. Some content switches now on the market support both URL switching and HTTP 1.1, but they route all requests on each connection to the server selected to handle the first request arriving on the connection. This sacrifices any cache affinity benefit from URL switching. Aron et al. [4], in a follow on to the LARD study [16] discuss this problem in detail and propose a solution using server to server transfers of requested data among the back ends. This approach is similar to cooperative caching [3, 12] of content objects among the servers, a solution found to be inferior in ....

M. Aron, P. Druschel, and W. Zwaenepoel. Efficient support for P-HTTP in cluster-based Web servers. In In Proceedings of USENIX'99 Technical Conference, 1999.

....in an untrusted environment by creating DVDs and managing them through a network virtual device manager. The proposed third party transfer scheme using DVDs is similar to that in NASD. The Linux NBD that is used in our prototype is similar to their virtual Internet SCSI adapter [17] In LARD [20, 5], outgoing data are directly sent to the clients by the back end server without going through the front end server through a TCP handoff protocol. A TCP handoff protocol is used to achieve client side transparency. After a TCP connection is handed off to a backend server, the front end server ....

M. Aron, P. Druschel, and W. Zwaenepoel. Efficient support for P-HTTP in cluster-based web servers. In Proceeding of the 1999.

....uses full process context checkpoints to recover from a crash and works only for single process servers. Compared to FT TCP, SC allows dynamic and fine grained connection migration of client sessions with state distributed over multiple communicating processes. TCP connection handoff is used in [3] for load balancing in clustered HTTP servers by distributing incoming client requests from a front end host to back end server nodes. Migration is limited to the initial connect. Multiple handoffs of persistent HTTP 1.1 connections at request granularity are mentioned, but no design or ....

M. Aron, P. Druschel, and W. Zwaenepoel. Efficient Support for P-HTTP in Cluster-Based Web Servers. In Proc. USENIX '99, 1999.

....15] They typically rely on either a proxy (which is a single point of failure) or maintaining a full connection state at the server client. In either case, they are not designed to tolerate server failures. Finally, several approaches have been taken to provide continued service of HTTP servers [2, 14]. 3 Informal Overview ST TCP is based on the idea of primary backup approach to provide high availability and fault tolerance. To ensure fast failover, ST TCP maintains active backup servers that can take over the functions of the primary server as soon as any failure of the primary server is ....

M. Aron, P. Druschel, and W. Zwaenepoel. Efficient support for P-HTTP in cluster-based web servers. In Proceedings of USENIX'99, 1999.

....one needs to decouple request dispatching from TCP connection set up. Previous systems used either TCP connection splicing [11, 30] which requires the front end to carry the splicing processing load in both incoming outgoing trac and thus tends to become the system bottleneck, or TCP hand o [26, 4, 5], which uses a connection state migration mechanism to move the TCP connection state either from the front end one to a back end node or from one back end node to another. Gage enjoys the scalability bene t of TCP hand o and the implementation simplicity of TCP splicing. Gage s TCP splicing is ....

M. Aron, P. Druschel, and W. Zwaenepoel. \Ecient Support for P-HTTP in Cluster-based Web Servers." In Proceedings of the USENIX 1999.

....mobile TCP might improve client perceived performance in face of network failures by using potential alternate routes to a server from a new location. Unlike M TCP, none of these schemes consider the task of migrating connection endpoints between physically distinct machines. TCP hando is used in [5] in clustered HTTP servers for load balancing by distributing incoming client requests from a front end host to back end server nodes. The system employs a limited, single hando scheme, in which a connection endpoint can only migrate during the connection setup phase. Multiple hando s of ....

M. Aron, P. Druschel, and W. Zwaenepoel. Ecient Support for P-HTTP in Cluster-Based Web Servers. In Proc. USENIX '99, 1999.

....(data transmission and ack receiving) are completed. However, this approach wastes network bandwidth by deliberately draining the data pipe before handing off the connection. In fact, Aron et al. have identified pipe draining as a potential problem that a connection handoff protocol must address [11]. Half pipe anchoring prevents data pipe drains during connection handoff. As soon as the optimal server for finishes sending the last data packet, it informs DServ that it is done. At this point, while the data packets from s optimal server are still unacknowledged, DServ can direct a ....

M. Aron, P. Druschel, and W. Zwaenepoel. Efficient support for P-HTTP in cluster-based web servers. In Usenix Annual Technical Conference, June 1999.

....serve the request. Since HTTP request is sent after the TCP connection is established, the dispatcher has to first establish a connection with the client before any distribution decision can be made. After that, the processing of the request is passed to the chosen web server node. TCP handoff [6, 7, 13, 25, 27, 31] and TCP splicing [15, 28] are two mechanisms that support such operation. In TCP handoff, the connection endpoint in the dispatcher is passed to the chosen node, which then processes the request and sends out the response directly to the client. In TCP slicing, the dispatcher acts like a proxy. ....

....the web server nodes, this in turn imposes more work in each of them. The overall performance is thus limited. Supporting persistent HTTP is another challenge in TCP handoff. The main issue is to prevent the TCP stream from draining during the process of handoff. This problem is mentioned in [6] but there is no solution given. The lack of such mechanism limits TCP handoff s advantage over layer 4 dispatching in persistent HTTP connections because only the first request in a connection can be dispatched according to its content. Subsequent requests in the same connection are still bound ....

[Article contains additional citation context not shown here]

M. Aron, P. Druschel, and W. Zwaenepoel. Efficient support for p-http in cluster-based web servers. In Proceedings of the USENIX 1999.

....[7] or a cluster based Web server. A cluster based Web server system consists of multiple Web servers connected by a high speed LAN. The Web servers work cooperatively to handle the Web requests. Many research and industry projects have been conducted on the design of cluster based Web servers [8,9,10,11 ], which aimed at such issues as load distribution or balancing [ 12,13,14] scalability [ 15,16] and high availability [ 17, 18] These projects have shown that clustering is a promising approach to building scalable and high performance Web servers [8,19] One of the key techniques to increase ....

....of documents. Furthermore, frequently redirecting client requests to a new server will result in constant TCP connection setups and breakdowns, making it difficult to implement HTTP 1.1 persistent connections to improve the system performance. 28 LARD Location Aware Request Distribution (LARD) [9,56] is a mechanism to distribute requests among backend server nodes in a cluster environment. The prototype system uses a frontend server as a request dispatcher. The dispatcher maintains a location mapping of objects to backend servers. Incoming requests will be distributed to the backend server ....

Aron M, Druschel P, Zwaenepoel W. Efficient Support for P-HTTP in Cluster- Based Web Servers. Proceedings of the 1999 USENIX Annual Technical Conference, Monterey, CA, June 1999.

....on providing high availability to services. This is so because the time to recover a particular connection can be high as reincarnated connections need to start from the beginning. 2.1. 4 TCP Connection Handoff in Cluster based Web Servers The Location Aware Request Distribution (LARD) project [5] proposes TCP connection handoff in clustered HTTP servers for distributing incoming requests from a front end machine to server back end nodes. It is limited to a single hand off scheme, where a connection endpoint can only migrate during the connection setup phase, whereas in our solution ....

M. Aron, P. Druschel, W. Zwaenepoel. Efficient Support for P-HTTP in Cluster-Based Web Servers. USENIX '99.

....is capable of serving requests independently. For NADs, such a scheme requires porting file system and applications (such as NFS HTTP server) on the disks, which essentially turning each disk into a small server. An extension of LARD which supports persistent HTTP connection is presented in [3]. A back end forwarding mechanism which forwards replies among the back end nodes is proposed to avoid the complexity and overhead of multiple TCP handoffs. In Slice [2] a request switching filter interposed along the network path between the client and the storage server routes file requests ....

M. Aron, P. Druschel, W. Zwaenepoel, and c Support. Efficient support for p-http in cluster-based web servers. In Proceeding of the

....most popular objects receive the largest number of requests, and a single server handles all requests for any given object. slb policies balance load, but they tend to scatter requests for each object across the servers, reducing server memory hits for moderately popular objects. Recent research [15, 2, 3] has studied this tradeoff in depth, and developed Locality Aware Request Distribution (lard) and related policies to balance these competing goals, combining the benefits of each approach. The lard approach is commercialized by Zeus technology [1] Other commercial request distributors use less ....

M. Aron, P. Druschel, and W. Zwaenepoel. Efficient support for P-HTTP in cluster-based Web servers. In In Proceedings of USENIX'99 Technical Conference, 1999.

....and request routing inside a single web server cluster. Several works analyze content based request distribution policies [PAB 98] ZBC 99] Others study mechanisms for redirecting requests to a particular web server, including DNS based [Bri95] TCP splicing [DKM96] CRS99] YL99] TCP hand off [ADZ99] [PAB 98] and distributed redirection [ASD 00] There are multiple studies that analyze the performance of web servers and the effects of different design choices. Pai, et al. looked into performance of different server architectures in context of one implementation: the Flash web server ....

M. Aron, P. Druschel, and W. Zwaenepoel. Efficient Support for P-HTTP in Cluster-based Web Servers. In Proceedings of the

....of persistent connections. Luo and Yang [7] augment server clusters with mechanisms that enable important sessions to be smoothly migrated and recovered on another node in the case of node failures. Their work is similar in the respect that they also identify important connections. Aron et al. [3] study mechanisms for supporting persistent connections in cluster based web servers that employ content based request distribution. Their aim is to improve throughput while we aim at preventing server overload. 3 Problem Description To prevent server overload, the amount of work entering the ....

....signal the web server that it can tear down the connection after sending the response. On the persistent connection, sclient sends request number N after the response for request number N 1 is received. Aron et al. have extended the sclient program in a similar way to handle persistent connections [3]. In our experiments each sclient program represents a group of clients requesting the same set of les. For simplicity, we will use the term sclient to denote the group of clients represented by one sclient program. The experimental workload consists of static and dynamic requests. The dynamic ....

M. Aron, P. Druschel, and W. Zwaenepoel. Ecient support for p-http in cluster-based web servers. In Usenix Annual Technical Conference, June 1999.

....connections. Bhatti et al. 9] have studied users tolerance of QoS in the context of e commerce. They found that users tolerance changes during the duration of a session. To enable the web server to keep track of session duration they propose to encode session duration in cookies. Aron et al. [5] study mechanisms for supporting persistent connections in cluster based web 2 servers that employ content based request distribution. Their goal is di erent from ours; they aim to improve throughput while we aim at preventing server overload. Banga and Druschel s [8] resource containers enable ....

....the response. On the persistent connection, sclient sends request number N after the response for request number N 1 is received. Aron et al. have extended the original sclient program in a similar way to evaluate their approach for supporting persistent connections in cluster based web servers [5]. In our experiments each sclient program represents a group of clients requesting the same set of les. For simplicity, we will use the term client to denote the group of clients represented by one sclient program. The experimental workload consists of static and dynamic requests. The dynamic ....

M. Aron, P. Druschel, and W. Zwaenepoel. Ecient support for p-http in cluster-based web servers. In Proc. of Usenix Annual Technical Conference, June 1999.

....on the computing resources of popular services available on the Internet. Clusters of commodity computers are currently being used to provide the scalability required by several of these services. Several researchers and companies have concerned themselves with the cluster based servers, e.g. [29, 31, 30, 24, 2, 3, 37, 11, 5, 12]. There are two main classes of servers in terms of how the clients requests are distributed across the cluster: content oblivious and content aware servers. In a content oblivious server the request distribution is based solely on a load metric, which is usually the number of open connections ....

M. Aron, P. Druschel, and W. Zwaenepoel. Efficient Support for P-HTTP in Cluster-Based Web Servers. In Proceedings of the USENIX 99 Annual Technical Conference, Monterey, CA, June 1999.

....connections. Bhatti et al. 9] have studied users tolerance of QoS in the context of e commerce. They found that users tolerance changes during the duration of a session. To enable the web server to keep track of session duration they propose to encode session duration in cookies. Aron et al. [5] study mechanisms for supporting persistent connections in cluster based web servers that employ content based request distribution. Their goal is di erent from ours; they aim to improve throughput while we aim at preventing server overload. Banga and Druschel s [8] resource containers enable the ....

....the response. On the persistent connection, sclient sends request number N after the response for request number N 1 is received. Aron et al. have extended the original sclient program in a similar way to evaluate their approach for supporting persistent connections in cluster based web servers [5]. In our experiments each sclient program represents a group of clients requesting the same set of les. For simplicity, we will use the term client to denote the group of clients represented by one sclient program. The experimental workload consists of static and dynamic requests. The dynamic ....

M. Aron, P. Druschel, and W. Zwaenepoel. Ecient support for p-http in cluster-based web servers. In Usenix Annual Technical Conference, June 1999.

....content information aware distribution, byletting the Web switch examine the HTTP request and then route it to the target server. The selection mechanism (usually referred to as delayedbinding) can be based on the Web service content requested, as URL content, SSL identi ers, and cookies. In [5] there are many techniques to realize the dispatching granularity at the session level or at the single Web object request level. Scheduling algorithms deployed at layer 7 may use either client information (as session identi ers, le type, le size) or a combination of client and server state ....

....in Web cluster nodes. The principle of LARD is to direct all requests for a Web object to the same server node. This increases the likelihood to nd the requested object into the disk cache of the server node. We use the LARD version proposed in [24] with the multiple hand o mechanism de ned in [5] that works for the HTTP 1.1 protocol. LARD assigns all requests for a target le to the same node until it reaches a certain utilization threshold. At this point, the request is assigned to a lowly loaded node, if it exists, or to the least loaded node. To this purpose, LARD de nes two threshold ....

[Article contains additional citation context not shown here]

M. Aron, P. Druschel, W. Zwaenepoel, \Ecientsupport for P-HTTP in cluster-based Web servers", Proc. USENIX 1999,Monterey, CA, June 1999.

....be sent directly from the service node to the client; or (2) be staged at the node that originally received the request. The former approach is e#cient in that the (potentially large) reply does not have to be transferred between two cluster nodes, but requires a TCP connection hand o# mechanism [16, 2] to transfer the TCP state of the node that accepts a request to the node that actually services the request transparently to the client. Unfortunately however, operating systems currently do not provide TCP hando # mechanisms, so implementing hand o# requires modifications to the operating system ....

....this gap, in this paper we study this tradeo# in the context of an interesting . External Network Internal Network MEM P NIe NIi MEM P NIe NIi MEM P NIe NIi Figure 1: Typical modern cluster of workstations. class of content based network servers, the locality conscious servers [16, 2, 3, 7, 5], under di#erent inter node communication subsystems. Locality conscious servers are inspired by work on cooperative caching [10, 13, 12] These servers use the set of memories of the cluster as a single large cache and distribute requests for read mostly files based on cache locality. This ....

[Article contains additional citation context not shown here]

M. Aron, P. Druschel, and W. Zwaenepoel. E#cient Support for P-HTTP in Cluster-Based Web Servers. In Proceedings of USENIX'99 Technical Conference, June 1999.

....client side of the connection (instead of blindly handing the connection off to a splice) and watch for further object requests. When it receives a new request, the wedge would begin announcing it via the softstate distribution protocol. Using techniques from the LARD persistent connection handler [4], the splice can still perform fast relaying of the bulk data from the server to the client. An in server implementation of the softstate dissemination protocol would of course eliminate this need. 4.2 Socket interface In order for a new server to handle a migrated connection, the wedge must ....

....TCP splicing [7] and handing off connections directly to end machines within a Web server cluster are well known. Handoff mechanisms were first explored by Hunt, Nahum, and Tracey [13] and later implemented in the LARD (Locality Aware Request Distribution) system [17] LARD was recently extended [4] to support request handoff between backend servers for HTTP 1.1 persistent connections [10] A key component of the LARD implementation is a TCP handoff mechanism, which allows the front end load balancer to hand the connection off to back end servers after the load balancing decision has been ....

[Article contains additional citation context not shown here]

M. Aron, P. Druschel, and W. Zwaenepoel. Efficient support for P-HTTP in cluster-based web servers. In Proc. USENIX '99, June 1999.

No context found.

M. Aron, P. Druschel, and W. Zwaenepoel. Ecient support for P-HTTP in cluster-based web servers. In Proceedings of the USENIX 1999.

No context found.

M. Aron, P. Druschel, and W. Zwaenepoel. Ecient support for P-HTTP in cluster-based web servers. In Proceedings of the USENIX 1999.

No context found.

M. Aron, P. Druschel, and W. Zwaenepoel. Ecient support for p-http in cluster-based web servers. In Proceedings of the USENIX 1999.

No context found.

Mohit Aron, Peter Druschel, and Willy Zwaenepoel. E#cient support for p-http in cluster-based web servers. In Proceedings of the USENIX 1999.

No context found.

Mohit Aron, Peter Druschel, and Willy Zwaenepoel. Ecient support for P-HTTP in cluster-based web servers. In Proceedings of the 1999.

First 50 documents

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC