Load Balancing for Heterogeneous Web Servers

Transcription

1 Loa Balancing for Heterogeneous Web Servers Aam Piórkowski 1, Aleksaner Kempny 2, Arian Hajuk 1, an Jacek Strzelczyk 1 1 Department of Geoinfomatics an Applie Computer Science, AGH University of Science an Technology, Cracow, Polan {aam.piorkowski,jacek.strzelczyk}@agh.eu.pl 2 Ault Congenital an Valvular Heart Disease Center University of Muenster, Muenster, Germany [email protected] Abstract. A loa balancing issue for heterogeneous web servers is escribe in this article. The review of algorithms an solutions is shown. The selecte Internet service for on-line echocariography training is presente. The inepenence of simultaneous requests for this server is prove. Results of experimental tests are presente 3. Key wors: loa balancing, scalability, web server, minimum response time, throughput, on-line simulator 1 Introuction Moern web servers can hanle millions of queries, although the performance of a single noe is limite. Performance can be continuously increase, if the services are esigne so that they can be scale. The concept of scalability is closely relate to loa balancing. This technique has been use since the beginning of the first istribute systems, incluing rich client architecture. Most of the complex web systems use loa balancing to improve performance, availability an security [1 4]. 2 Loa Balancing in Cluster of web servers Clustering of web servers is a metho of constructing scalable Internet services. The basic iea behin the construction of such a service is to set the relay server 3 This is the accepte version of: Piorkowski, A., Kempny, A., Hajuk, A., Strzelczyk, J.: Loa Balancing for Heterogeneous Web Servers. In: Kwiecien, A., Gaj, P., Stera, P. (es.) CN CCIS, vol. 79, pp Springer, Heielberg (2010) The original publication is available on

2 2 Loa Balancing for Heterogeneous Web Servers in a transparent connection to the cluster servers (Fig. 1). This server is calle a ispatcher. Cluster of Web Servers INTERNET DISPATCHER Fig. 1. Cluster of web servers schema. There are many implementations of loa balancing. Some of them (like OpenSSI for Linux or MS Winows Network Loa Balancing) use aitional software that have to be installe on each cluster noe. This software monitors the loa of these noes, but is eicate for selecte operating systems or web server software. It makes a construction of heterogeneous web servers clusters impossible. This article focuses only on some implementations that allow to create heterogeneous web server clusters. This means that it is esirable to create services base on the noes that use ifferent operating systems an ifferent environments, but process the same requests. Technique which realizes loa balancing fulfilling these assumptions is proxy loa balancing. 2.1 The algorithms of loa balancing Efficient loa balancing requires an appropriate algorithm. There are several basic an common algorithms being iscusse further in this paper: Roun Robin [1], Weighte Roun Robin, Least Loae [1, 2], Least Connection [1], Weighte Least-Connection, Locality-Base Least-Connection, Destination Hashing [13], Source Hashing [12, 13],

3 Loa Balancing for Heterogeneous Web Servers 3 Fair [11], Never Queue, Shortest Queue First [14], Request Counting [9], Weighte Traffic Counting [9], Pening Request Counting [9]. The Roun-Robin algorithm is a well-known algorithm an it is easy to implement. Request Counting algorithm istributes the requests among the various noes to ensure that each noe gets its configure share of the number of requests [9]. Weighte Traffic Counting algorithm works in a similar way, but the maximum number of requests per server is etermine on the network traffic, in bytes. Pening Request Counting algorithm s iea is to keep track of how many requests each worker is assigne at the time. New requests are assigne to the server with the lowest number of active requests. The fair algorithm is base on the stanar roun-robin algorithm an it tracks busy back en servers an balances the loa to non-busy servers [11]. There are many other algorithms, some of them require special knowlege to preict the best scheuling [2]. 2.2 Solutions There are a few implementations of loa balancing proxies that enable to create a heterogeneous cluster of web servers. The most interesting from the perspective of authors of scientific portals (non-commercial) are open software solutions. In this article we iscusse six of them: Apache Server with Mo Proxy Balancer [9] - this is a stanar solution use in the most popular web server, it implements three loa balancing algorithms : Request Counting, Weighte Traffic Counting an Pening Request Counting, Poun [10] - a simple solution with Roun-Robin algorithm, istribute uner the GPL, NGiNX [11] - this software implements Roun Robin an Fair loa balancing algorithms, NGiNX is license uner 2-clause BSD-like license, Inlab Balance [12] - is an open source loa balancer, uner the GPL licensing terms, that implements two algorithms - Roun-Robin an Client Hashing, HAProxy [13] - this solution implements stanar Roun Robin algorithm an others - Source Hashing an Destination Hashing, Lighttp [14] - this is one of the famous an efficient web server, but also proxy balancer, that implements four algorithms: Static (fail-over), Roun Robin, Cache Array Routing Protocol (similar to Source Hashing) an Shortest Queue First. 3 Features of the web servers Performance is one of the most important aspects of scalable web services. To etermine the performance the following factors shoul be consiere:

4 4 Loa Balancing for Heterogeneous Web Servers average response time (t avg ), minimum response time (t min ), throughput (th). The average response time is a factor which value varies an epens on users loa. Its value increases with the number of users. The minimum response time is the minimum time in which a web request is complete. It epens on the server performance (harware, software) an the request type. It can be constant for the same conitions. It shoul be measure at minimal loa of a server. The throughput is a very authoritative factor that escribes the performance of a server. It tells how many requests can be processe at the unit of time at the saturation. However, the system that reache the maximum throughput cannot guarantee the acceptable response time. 3.1 The efficiency of request processing The requests that are processe by a server can be of two types: inepenent request, relate request. The inepenent requests are requests, that o not affect one another. They share resources (for example CPU), which are share fairly between them. At saturation of a web server with one processor the relationship of minimum response time an throughput for this case can be given by the efficiency factor (1): E 1 = t min th. (1) The efficiency factor E 1 for series of ientical inepenent request shoul have the value close to 1.0. For multiprocessor servers the value of E 1 factor shoul be close to the number of processors. In this case the efficiency factor is given by formula (2), where N - number of processors. E = t min th/n. (2) Another type of queries are the relate. Mechanisms of optimization like caching or spooling can make processing shorter for a group of requests. This is for example the case of queries with pool connections to atabases [5, 6]. In this case the value of efficiency factor is above the number of processors. Some queries can generate a large overhea (for example allocating an eallocating big ata tables, frequent context switching) - requests affect each other an the efficiency factor is below the number of processors.

5 Loa Balancing for Heterogeneous Web Servers CT2TEE - an example of a web server CT2TEE is a novel Internet-base simulator for transesophageal echocariography (TEE) training [7]. It loas CT ata into the memory an processes projections [8] an USG simulations for each request iniviually. The process of creating a projection is escribe on Fig. 2. CLIENT CT2TEE Server Loaing CT ata into memory Preprocessing Picture request Calculating the projection(s) Simulating USG artifacts Picture output Compression JPG/GIF Fig. 2. Diagram for a request processing by CT2TEE server. The output of CT2TEE application can be an image, that is a single frame (of JPG or GIF format, JPG quality: 80%) or an animation (of GIF format). The current version of CT2TEE generates the same projection with ifferent noise pattern, but there will be motion implemente in the future. The GIF format generates bigger files than JPG. The one of the most interesting features of CT2TEE application is a fact, that the efficiency factor (2) in this case on the current Internet server of CT2TEE (2 processors) is very close to value 1 (0.99). It is cause by the character of requests - they are calculations that share CPU only. Therefore the CT2TEE application is a goo example to test loa balancing on a cluster of servers. 4 Tests The tests have been carrie out to assess performance. 4.1 Harware an software environment The following harware has been use for the tests:

6 6 Loa Balancing for Heterogeneous Web Servers for web servers/proxy server/test clients: IBM Blae HS 21, CPU: 2.0 GHz, Intel Xeon (8 cores), RAM 16GB, network: Ethernet 1Gb, switch. The following software was use: operating systems: Linux Feora Core 12, Winows Server 2008, component environments: Mono ,.NET 2.0, web servers: Mono XSP 2.4.2, IIS 7.0, loa balancers: Apache Mo Proxy 2.2, NGiNX , Poun 2.5-1, Inlab Balance 3.52, HAProxy an Lighttp , loa testers: JMeter, Apache Bench. The results given by JMeter an Apache Bench were very similar, so we ecie to use JMeter in all cases. The tests were ivie into two parts: etermining iniviual parameters of servers, etermining performance of loa balancing. 4.2 The efficiency of servers Initially the tests for the main parameters of cluster servers have been carrie out. The results (minimum times of requests t min [ms], throughputs th [req/s] an efficiency factors E) are presente in Table I. Table 1. Minimum times of requests t min [ms], throughputs th [req/s] an efficiency E for cluster servers with CT2TEE application. OS server G0 G1 G4 t min th E t min th E t min th E Linux s XSP s s s WinSvr s IIS s s s The performance of loa balancing Experiments for the six solutions (Apache Mo Proxy, NGiNX, Poun, Inlab, HAProxy an Lighttp) were one. In the case of Apache Mo Proxy we teste all three algorithms: Request Counting (RC), Weighte Traffic Counting (WTC) an Pening Request Counting (PRC). In the case of Inlab we teste only the

7 Loa Balancing for Heterogeneous Web Servers 7 Roun Robin algorithm. In the case of HAProxy we teste three algorithms: Roun Robin (RR), Source Hashing (SRC) an Destination Hashing (URI). In the case of Lighttp we teste four algorithms: Cache Array Routing Protocol (CARP), Roun Robin (RR), Static (failover balancing, STAT) an Shortest Queue First (SQF). There were two kins of heterogeneous environments: 3 servers running Linux+XSP an 1 server running Winows+IIS, 1 server running Linux+XSP an 3 servers running Winows+IIS. We selecte output to be a JPG (G0, small files of average size 20 KB) an GIF (G1, bigger files of average size 80 KB for 1 frame an 300 KB for 4 frames - G4). The G4 output was processe much longer than others. The results are presente in table II an on the plots (Fig. 3, 4). To compare these results with the maximum performance of a system an aitional column (MAX) is place. It contains sums of all server throughputs for the teste cases. Table 2. The performance of loa balancing - throughputs [req/s]. Loa Balancer 1 WinIIS + 3 LinXSP 3 WinIIS + 1 LinXSP G0 G1 G4 G0 G1 G4 Poun NGiNX RR NGiNX FAIR Inlab Apache WTC Apache RC Apache PRC HAProxy RR HAProxy SRC HAProxy URI Lighttp CARP Lighttp RR Lighttp STAT Lighttp SQF MAX Summary The tests have prove that the use of proxy loa balancers effectively increases system throughput. Some of the servers provie several algorithms, the choice of one of them is crucial for performance. For teste solutions using the CT2TEE application server the best results are reache by Inlab, Lighttp with Shortest Queue First algorithm an Apache Mo Proxy with Pening Request Counting algorithm. Slightly smaller througputs were achieve for the other solutions

8 8 Loa Balancing for Heterogeneous Web Servers T h r o u g h p u t [r e q / s ] G0 G P oun x RR x FAIR I nlab A pac h ew T C A A pac pac h h erc epr C H H H L L L L [ APr APr APr i g MA htp o o o xy xys xyu ightp ightp ightp X IMU R CA RR STA SQ R C R M I R P T F ] P oun N N I gin gin nlab A A A H H H pac pac pac APr APr APr x x RR FAIR h h h ew erc epr o o o xy xys xyu T C C R R C R I 1WinIIS + 3LinXSP 3WinIIS + 1LinXSP L i g htp CA R P RR STA T SQ F [ MA X IMU M ] Fig. 3. The results of loa balancing tests for G0 (JPG) an G1 (GIF, 1 frame) output.

9 Loa Balancing for Heterogeneous Web Servers 9 T h r o u g h p u t [r e q / s ] P oun x RR x FAIR I nlab A pac h ew T C A pac h erc A pac h epr C H APr o xy R H APr o xys R C H APr o xyu R I CA R P RR L i g htp STA T L i g htp SQ F [ MA X I M U M ] P oun x RR x FAIR I nlab A pac h ew T C A pac h erc A pac h epr C H APr o xy R H APr o xys R C H APr o xyu R I 1WinIIS + 3LinXSP 3WinIIS + 1LinXSP CA R P RR L i g htp STA T L i g htp SQ F [ MA X I M U M ] Fig. 4. The results of loa balancing tests for G4 (an animate GIF, 4 frames) output.

10 10 Loa Balancing for Heterogeneous Web Servers that use Roun Robin algorithm - NGiNX (Roun Robin an Fair algorithms), Poun, HAProxy (with Roun Robin algorithm) an Lighttp (with Roun Robin algorithm). The worst results were prouce by proxy balancers with Source Hashing an Destination Hashing algorithms. Apache Mo Proxy with Weighte Traffic Counting algorithm is over two times slower than the best results, but this algorithm is better in case of variable size of outputs. As we prove the choice of solution an algorithm is very important to reach the maximum performance of web server clusters. Acknowlegment. This work was finance by the AGH - University of Science an Technology, Faculty of Geology, Geophysics an Environmental Protection as a part of statutory project number References 1. Teo, Y.M. an Ayani, R., Comparison of loa balancing strategies on cluster-base web servers. Simulation. vol. 77 issue 6. p (2001). 2. Guo J., Bhuyan L. N.: Loa Balancing in a Cluster-Base Web Server for Multimeia Applications. IEEE Transactions On Parallel An Distribute Systems, vol. 17, no 11, (2006). 3. Ungureanu V., Melame B. an Katehakis M.: Effective loa balancing for clusterbase servers employing job preemption. Performance Evaluation, vol. 65, issue 8, p (2008). 4. Wrzuszczak, J.: Auction mechanism in management of processing noes in a computer cluster. Contemporary Aspects of Computer Networks 2, (2008). 5. Bogari-Meszoly A., Szitas Z., Levenovszky T., Charaf H.: Investigating Factors Influencing the Response Time in ASP.NET Web Applications. Lecture Notes in Computer Science (LNCS), Springer, vol , pp (2005). 6. Gabiga A., Piórkowski A., Danek T.: Efficiency analysis of servlet technology in selecte atabase operations. Stuia Informatica, nr 84, vol 30 issue 2B, (2009). 7. Kempny A., Piórkowski A.: CT2TEE - a Novel, Internet-Base Simulator of Transoesophageal Echocariography in Congenital Heart Disease. Kariol Pol 2010; 68: (2010). 8. Piorkowski A., Jajesnica L., Szostek K.: Creating 3D Web-Base Viewing Services for DICOM Images. Computer Networks, 16th Conference, CN 2009, Wisla, Polan, June 16-20, 2009, Communications in Computer an Information Science, Springer Berlin (2009). 9. Mo Proxy Balancer - Apache HTTP Server, Poun - Reverse-Proxy an Loa-Balancer, NGiNX - HTTP an reverse proxy server, Inlab Balance, HAProxy - The Reliable, High Performance TCP/HTTP Loa Balancer, Lighttp - fly light,