G. Banga and J.C. Mogul. Scalable kernel performance for Internet servers under realistic loads. In Conference, New Orleans, LA, 1998. 13  Home/Search   Document Details and Download   Summary   Related Articles   Check

.... number of simultaneously open connections (a typical value of 256) but newer systems are shipped with higher maximum values (typically thousands) Studies show, however, that with current implementations, large numbers of (even idle) connections can have a detrimental impact on server s throughput [2, 3, 4]. The design challenge of good connection management policies is to strike a good balance between benefit and cost of maintaining open connections and to enforce some quality of service and fairness issues. The current version 1.3 of the Apache HTTP Server [1] uses a fixed holding time for all ....

....significant reductions, typically around 50 , are achieved on the clicks logs. The resulting performance improvement is considerable, since connection establishments induce user perceived latency and overhead [16, 31, 23] wheras large number of open connections is both detrimental to throughput [4] and is more likely to reach the server s hard set limits, causing it to refuse new connections. In Section 2 we discuss the interaction between HTTP and TCP at Web servers, and the nature of server logs data. We then describe the logs used in our evaluation. Last, we discuss issues with the use ....

[Article contains additional citation context not shown here]

G. Banga and J. Mogul. Scalable kernel performance for Internet servers under realistic loads. In Proceedings of the USENIX Annual Technical Conference. USENIX Association, 1998. http://www.cs.rice.edu/~gaurav/papers/usenix98.ps.

....The studies on improving Web server s macroperformance focus on improvement of either the interactivity between the Web servers, the underlying operating systems, or disk I O and network I O. The studies in this field generally fall into three categories: operating system enhancement [2,6,7,8], server program improvement [1,5] and caching techniques [12,15] A very limited number of studies focus on architectural performance of Web servers, which we call micro performance in this paper. Radhakrishnan and John [23] evaluated the performance of Apache Web server in terms of ....

G. Banga, J.C. Mogul, Scalable Kernel Performance for Internet Servers Under Realistics Loads, Proceedings of 1998.

....are available to a user space web server without modifying the operating system. It caches les in user space with mmap( caches stat( information, caches URI lookups, and exploits writev( Flash Poll is a version of Flash which we modi ed to use poll( rather than select( Banga and Mogul [4] have shown that event driven servers that use select( such as Flash, can su er performance problems when managing large numbers of active connections. To avoid these problems, we adapted Flash to use poll( a similar function derived from SVR4, which is available on some forms of Unix, ....

Gaurav Banga and Jerey C. Mogul. Scalable kernel performance for Internet servers under realistic loads. In USENIX Annual Technical Conference, New Orleans, Louisiana, June 1998.

....with mmap( caches stat( information and URI lookups, and exploits writev( Flash has been shown to be significantly faster than Apache. We extended Flash in several ways to obtain the maximum performance in our system [20] First, we modified it to poll( rather than select( Banga and Mogul [5] have shown that event driven servers that use select( such as Flash, can suffer performance problems when managing large numbers of active connections. To avoid these problems, we adapted Flash to use poll( a similar function derived from SVR4, which is available on some forms of Unix, ....

Gaurav Banga and Jeffrey C. Mogul. Scalable kernel performance for Internet servers under realistic loads. In USENIX Annual Technical Conference, New Orleans, Louisiana, June 1998.

....instance, it is not clear under which circumstances a policy should delay a request before the locality benefit disappears. Thread scalability is limited when building highly concurrent applications [Ous96] PDZ99] Related work suggests inexpensive implementations for context switching [AB 91][BM98], and also proposes event driven architectures with limited thread usage, mainly for internet services [PDZ99] Welsh et al. propose a staged event driven architecture (SEDA) for deploying highly concurrent internet services [WCB01] SEDA decomposes an event driven application into stages ....

G. Banga and J. C. Mogul. "Scalable kernel performance for Internet servers under realistic loads." In Proc. USENIX, 1998.

....are poll( and select( However, neither of these calls scale very well as the number of descriptors being monitored for events becomes large. A high volume server that intends to handle several thousand descriptors quickly finds these calls becoming a bottleneck, leading to poor performance [1] [2] 10] The set of events that the application may be interested in is not limited to activity on an open file descriptor. An application may also want to know when an asynchronous I O (aio) request completes, when a signal is delivered to the application, when a file in the filesystem changes ....

BANGA, G., AND MOGUL, J. C. Scalable kernel performance for Internet servers under realistic loads. In Proceedings of the 1998.

....event handling code to be short and run to completion, to avoid stalling the event processing threads, and to ensure fairness across a large number of requests. This typically requires that nonblocking, asynchronous I O interfaces be used. While much prior work has explored scalable I O interfaces [12, 13, 84, 111, 113], and nonblocking socket interfaces are common, many operating systems do not provide true nonblocking I O for disk access. Flash, Harvest, and other systems address this challenge using separate threads or processes to perform long running operations asynchronously. However, it is also possible ....

G. Banga and J. C. Mogul. Scalable kernel performance for Internet servers under realistic loads. In Proceedings of the 1998.

....the size of the requested file across runs. Note that we do not use a more realistic request pattern because our goal is to stress different aspects of the networking stack, and constant file sizes are more useful for this purpose. Shorter file sizes place more emphasis on per connection overheads [6], for example, SILK path creation and destruction. On the other hand, inefficiencies in the SILK protocol stack, such as excessive data copying, should be more visible using large files. For this experiment we configure all SILK paths with equal priorities. Since the SILK kernel task runs at a ....

G. Banga and J. Mogul. Scalable kernel performance for internet servers under realistic loads. In Proc. of Usenix Annual Technical Conference, New Orleans, LA, USA, June 1998.

....characterizes the Web (the number of online users is growing at about 90 per annum) Many efforts have also been directed at improving the performance of a Web server node at the software level, namely software scale up. This includes the operating system and also the Web server application [14, 12, 57, 75, 80]. The Flash Web server ensures that its threads and processes are never blocked by using an asymmetric multiprocess event driven architecture [80] Nahum et al. 75] have analyzed how a general purpose operating system and the network protocol stack can be improved to provide support for ....

G. Banga and J. C. Mogul. Scalable kernel performance for Internet servers under realistic loads. In Proc. of USENIX 1998.

....the ability to multiplex between outstanding connections. Significant research has been conducted into improving web server performance by improving both operating system mechanisms and interfaces for obtaining information about the state of socket and file descriptors from the operating system [3] [12] 2] 4] 13] 14] 6] These studies have been motivated by the belief that under high loads with a large number of concurrent connections, the overhead incurred by select (or similar calls) is prohibitive to implementing high performance Internet servers. As a result they have mainly ....

G. Banga and J.C. Mogul. Scalable kernel performance for Internet servers under realistic loads. In Proceedings of the 1998.

....is needed to deliver these events from multiple endpoints, and allow efficient synchronisation with event arrival. Traditionally some variant on the BSD select system call is used. However, the deficiencies of this style of interface are well known, and even with sophisticated optimisations[8] have been shown to incur high overhead and scale poorly with the number of endpoints. We have developed a mechanism which delivers events directly to the application with low overhead, and with cost O(1) using the asynchronous queue described above. The queue resides in memory mapped into the OS ....

Gaurav Banga and Jeffrey Mogul. Scalable kernel performance for Internet servers under realistic loads. In USENIX Technical Conference, June 1998.

.... web server performance [11] Likewise, there has been considerable prior work in performance analysis and benchmarking of conventional web servers [15, 12, 17, 5, 18] performance optimizations of web servers, performance oriented web server design, and operating system support for web servers [13, 22, 6, 7, 21]. Apostolopuolos et al. 3] studied the cost of TLS connection setup, RC4 and MD5, and proposed TLS connection setup protocol changes. Our methodology is to replace each individual operation within TLS with a no op and measure the incremental improvement in server throughput. This methodology ....

G. Banga and J. C. Mogul. Scalable kernel performance for Internet servers under realistic loads. In Proceedings of the

....of a web accelerator deployed as the front end to a set of web servers [LIS 99] Several works analyze the cost of different operations of a busy web server. Some of these studies use a simulated workload [HMS98] BC98] while others study server behavior under reallife workload [MRG97] CDN 96] [BM98]. These studies show the importance of reproducing such workload characteristics as request inter arrival time, network delays, and a large number of simultaneously open connections. There are several areas in operating systems design that are important for web server performance. U Net proposed ....

....through them. Such set up avoids connection open and close overhead, which is usually incurred when clients send requests directly to the accelerator. Also smaller number of concurrent connections decreases the cost of packet demultiplexing, which may significantly affect the overall performance [BM98]. 7 Device Driver TCP IP HTTP Server Socket Accelerator Socket read Packet Cache Packet send File System Figure 3. Accelerator design The web booster delays acknowledgements in order to decrease the overhead of network packet processing. When a booster and an origin server are ....

G. Banga and J. C. Mogul. Scalable Kernel Performance for Internet Servers Under Realistic Loads. In Proceedings of the USENIX Annual Technical Conference, New Orleans, LA, June 1998.

....Section 5 summarizes and concludes. 2 Related Work Previous work in the performance analysis and improvement of Web transactions falls into two categories: work on servers, and work on networks protocols. Work on servers has emphasized techniques for improving their performance, for example [2, 6, 8]. Such studies show how Web servers behave under a range of loads (including overload conditions) and have suggested enhancements to application implementations and the operating systems on which the servers run. However, unlike this paper, such studies have not to date considered how interactions ....

G. Banga and J. Mogul. Scalable kernel performance for internet servers under realistic loads. In Proceedings of the USENIX Annual Technical Conference,

.... contrast to our study which investigates the tail of the load distribution) Finally, other issues such as strong and weak consistency mechanisms [6, 17, 32] cache replacement algorithms [5, 20] hierarchical proxy caching [29, 31, 33] prefetching techniques [14, 21, 25, 34] and protocol issues [3, 24] have also received significant research attention; these efforts are complementary to our current study. 6 Concluding Remarks In this paper, we examined the potential benefits of web proxy caches in improving the effective capacity of servers and networks. Since servers and networks are ....

G Banga and J C. Mogul. Scalable Kernel Performance for Internet Servers under Reslistic Loads. In Proceedings of the USENIX Annual Conference, New Orleans, LA, June 1998.

....patterns affect Web performance is limited. Because of the variability of real world studies and lack of reproducibility, controlled laboratory studies using synthetic workloads remain important. A number of studies have characterized Web server performance using synthetically generated workloads [17, 7, 14, 15, 78, 100]. These studies have pointed out that server bottlenecks can occur in a subsystems such as CPU or disk, and that inefficiencies in specific operating system calls can limit performance. However, unlike our work, none of these studies have considered how interactions of networks, clients and ....

....of networks, clients and servers in the wide area affect end to end performance. In [7] Almeida et al. evaluate the burden of lingering TCP connections in HTTP 1.0. They also show that for the workloads they consider, the server s CPU is typically the bottleneck in the system. The work in [15] contains a detailed analysis of a server under heavily loaded conditions. Through kernel profiling, Banga and Mogul find inefficiencies in the select and ufalloc system calls in Digital s UNIX. Improvements in the implementations of these calls lead to significant improvements in overall ....

[Article contains additional citation context not shown here]

G. Banga and J. Mogul. Scalable kernel performance for Internet servers under realistic loads. In Proceedings of the USENIX '98, New Orleans, LA, June 1998.

....or may not be time critical. Event noti cation services such as the BSD select system call do not cope very well with these requirements. On the one hand, the original BSD implementation does not scale very well with the number of event sources and is ill suited for time critical event delivery [BM98] A more recent implementation scales better but still consumes a high amount of CPU time because of properties inherent to the semantics of select [BMD99] On the other hand, while many event types in our multimedia application originate within bu er and con guration managers, select is de ned ....

Gaurav Banga and Jerey C. Mogul. Scalable Kernel Performance for Internet Servers Under realistic Loads. In Proceedings of the

....range in load, with the latency of each task increasing linearly. An important limitation of this model is that it assumes that eventhandling threads do not block, and for this reason nonblocking I O mechanisms must be employed. Although much prior work has investigated scalable I O primitives [8, 9, 33, 46, 48], event processing threads can block regardless of the I O mechanisms used, due to interrupts, page faults, or garbage collection. Event driven design raises a number of additional challenges for the application developer. Scheduling and ordering of events is probably the most important concern: ....

G. Banga and J. C. Mogul. Scalable kernel performance for Internet servers under realistic loads. In Proc. 1998 Annual Usenix Technical Conference, New Orleans, LA, June 1998.

.... and the kinds of events of interest to them (events providing information of activity on TCP IP sockets) Such web servers present a natural opportunity to evaluate our mechanism and despite the signi cant interest within the research community that this speci c problem domain has generated [9, 10, 27, 18], we still see room for improvement. 1.2.1 Multiplexing Connections in Web Servers The Internet has seen phenomenal growth since its creation. This growth has seen systems initially designed to work in small academic or corporate environments being pressed into use in a world wide public network ....

....of today s Internet servers run on UNIX [24] an operating system originating in the 1970 s. There is evidence to show that the mechanisms and interfaces of UNIX systems for managing multiple connections scale poorly with respect to the number of clients served by Internet server applications [9, 10, 27]. Web servers are essential to the World Wide Web [12] their role being to service the requests of remote clients using the Hyper Text Transfer Protocol (HTTP) 16] Each client conducts an HTTP dialog with the server, requesting and receiving data over a TCP IP connection. Each TCP IP ....

[Article contains additional citation context not shown here]

Gaurav Banga and Jeery C. Mogul. Scalable Kernel Performance for Internet Servers Under Realistic Loads. In Proceedings of the USENIX 1998 Annual Technical Conference, 1998.

....table, and then calls vn open to perform the open. Times for these routines are shown in Figure 6.2. falloc takes negligible time, a surprising finding because Squid file table allocation has been reported to be a bottleneck under heavy load in a previous study (performed on Digital UNIX) [8]. Since copen spent most of its time in vn open, the performance of vn open was examined next. Here, file creation diverges from opening an existing file; vn create is called if the O CREAT flag was Figure 6.1 copen. Although called only 20 25 times sec, copen is a clear bottleneck. On average, ....

G. Banga and J.C. Mogul. Scalable Kernel Performance for Internet Servers Under Realistic Loads. 1998 USENIX Annual Technical Conference, New Orleans, LA, June 1998.

No context found.

G. Banga and J.C. Mogul. Scalable kernel performance for Internet servers under realistic loads. In Conference, New Orleans, LA, 1998. 13

No context found.

Gaurav Banga and Jeffrey C. Mogul. Scalable kernel performance for Internet servers under realistic loads. In Proceedings of the 1998.

No context found.

G. Banga and J.C. Mogul. Scalable kernel performance for Internet servers under realistic loads. In Proceedings of the USENIX 1998 Annual Technical Conference, pages 1--12, 1998.

No context found.

G. Banga and J. Mogul. Scalable Kernel Performance for Internet Servers Under Realistic Loads. In Proceedings of the USENIX Technical Conference, June 1998.

No context found.

Gaurav Banga and Je#rey Mogul. Scalable kernel performance for Internet servers under realistic loads. In USENIX Technical Conference, June 1998.

First 50 documents  Next 50

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