Experimental Study of Parallel Downloading Schemes for Internet Mirror Sites

Transcription

1 Experimental Study of Parallel Downloading Schemes for Internet Mirror Sites Spiro Philopoulos Dept. of Elect. and Comp. Eng. University of Manitoba Winnipeg, Canada Muthucumaru Maheswaran Dept. of Computer Science University of Manitoba Winnipeg, Canada ABSTRACT A common method used to reduce document retrieval times is the use of content replication i.e., mirror servers. The mirror servers provide several alternate sites to download a specific document and were traditionally used to increase the availability of content. Recently, several studies focused on using multiple mirror sites to concurrently download portions of a document from a set of mirror sites. Following are some of the issues involved in using multiple mirror sites concurrently: (a) selection of the best mirror servers from the client, (b) coping with dynamic overloading of the network and servers, and (c) coping with faults. This paper briefly examines two existing schemes for concurrent downloading or parallel-access downloading, or paraloading as it is called. It proposes a third paraloading scheme called the Dynamic Parallel Access. The performance of this scheme is experimentally evaluated. Recommendations for further improvements are also discussed. 1. Introduction As various applications are Internet-enabled, the number of repositories in the Internet that hold valuable content are increasing. For example Internet-based repositories are beginning to hold content such as multi-gigabit movie files, operating system and other large software distributions, and large multimedia documents. This creates a need for the clients to find faster downloading schemes for both online and offline usage. The conventional way of downloading files from an Internet-based server is to open one or more connections between it and the client. While opening multiple connections might reduce download times compared to a single connection, due to the following issues the performance enhancement can be limited: (a) server load and capacity, (b) bottleneck link bandwidth, (c) instantaneous bandwidth, (d) multi-connection client overhead, and (e) interconnection resource allocation. Replicating content such that it can be accessed from multiple locations is one way for decreasing the download times. Caching, content delivery, and mirroring are some of the techniques that replicate content using different policies and primary purposes. Mirror servers were traditionally used to improve the availability of the content. Recently, however, several projects have examined the concept of concurrently using multiple mirror servers to download content for a single client. The concurrent or parallel downloading is performed to reduce the download time for a client. One of the key problems with paraloading schemes is the mirror site selection problem. This problem is complicated because accurate performance information to select the best set of mirror servers is not available to the clients. Additionally, the problems mentioned above for the single server case still exist. Further, network conditions can change during the download that can lead to decreased performance. This observation motivated us to examine adaptive parallel download techniques that use a set of mirror servers. The membership of the mirror server set may dynamically change. In paraloading segments of the file are downloaded from each server in the set and are then reassembled at the client. The parallel-access scheme was first proposed by [4]. Following are some of the advantages of paraloading: Depending on the topology, in the ideal case, the aggregate bandwidth of all individual connections may increase the overall throughput to the client. Because multiple connections are used, paraloading is more resilient to link and general route failures. The inherent load balancing associated with parallel loading due to the fact that connections are spread out to many servers and not just one. Therefore, the scheme can be immune to individual server load fluctuations, bottleneck link bandwidth and traffic fluctuations. Section 2. examines three different paraloading schemes (two existing and one new) all of which use application-level negotiation to schedule the transmission of different segments of a file. Section 3. describes the experiments performed to evaluate the different schemes and examines the results obtained from the experiments.

2 2. Parallel Downloading Schemes 2.1 History-Based Parallel Access History-based parallel access [4] is a relatively simple scheme in which information regarding the previous transmission rates between the client and every mirror server is recorded in a database and used to determine how large a file segment is downloaded by each server. More specifically, for Å mirror servers, the file will be divided into Å unequally-sized disjoint segments each of which will be assigned to a mirror server based on expected data rates derived from historical data between the client and each mirror, i.e., the faster a server is (based on the history) the larger the file block it will be assigned to be downloaded from that server. File blocks are downloaded from each individual server using the HTTP 1.1 byte range header feature resulting in a server-side transparent solution requiring modifications only on the client side. The problem with the history-based parallel access is of course the validity of the recorded transmission rates i.e., how close those recorded rates are compared to the actual transmission rates in future downloads. The larger the divergence between historical and (future) actual data transmission rates, the worse the performance is since slower servers will be assigned file blocks larger than should have been and vice versa for mirror servers faster than what historical data proclaims. One issue that must be addressed with such a scheme is how exactly is server transmission rate data obtained and updated. 2.2 Semi-dynamic Parallel Access The semi-dynamic parallel access downloading method was first proposed by [4], where it is referred to as dynamic parallel access downloading but in this paper is referred to as semi-dynamic in order not to confuse with our own parallel downloading method examined next. The semidynamic scheme is conceptually simple: initially the receiver will obtain the size of the file it wishes to download (e.g., by polling one of the mirror servers) and then the file is segmented by the receiver into equal-size blocks. The receiver will request at the beginning from all the mirror servers to download one block. Once a server has completed sending the requested file block, the client will request a new (undelivered) block from the server. This will continue until all file blocks have been downloaded, at which point the receiver will reassemble the file from the individual downloaded blocks. As in history-based parallel access downloading, the HTTP 1.1 byte range feature is used to download individual blocks of the file from each mirror server. In order to reduce overhead, persistent TCP connections are used between the client and each server. One enhancement used is that if there are less file blocks left than servers, a file block that has already been requested but not yet completely downloaded can be downloaded simultaneously in parallel by another server leading potentially to a faster download for that block. In this scheme, faster servers provide larger portions (more blocks) of the total transmitted data. Based on the paraloading scheme proposed in [4], [3] proposed a modified paraloading scheme which is essentially identical to that proposed by [4] but with the following three enhancements: Minimizing the delay at startup by piggybacking onto a data block a request for the file size and a request for the list of mirror servers that posses the requested file. Minimizing the idle time between block downloads by pipelining the block requests. Minimizing the idle time in downloading the last block in one of three ways: a) use small block sizes b) dynamically adjust file block size for that last blocks c) send requests to the idle servers to download the remaining portions of the last block. 2.3 Dynamic Parallel Access In this section a new type of paraloading scheme developed by the authors of this paper is examined, called dynamic parallel access downloading. This new paraloading method is based to a certain extent on the semi-dynamic paraloading scheme first proposed by [4], as examined above, however it was designed with large size file transfers in mind. In this new scheme, like those before, the client segments the file into fixed-size blocks which are requested and downloaded from the individual mirror servers. More specifically, operation is as follows. Initially connections are opened to all the mirror servers. In the current implementation the list of available mirror servers is obtained from a file containing the server list, however in the future for actual practical use there must be a way to dynamically obtain the list of available servers. This could be done in various ways, such as retrieving the server list from some type of a directory service, or by extending the DNS system to provide such information [2]. In order to reduce overhead, persistent TCP connections are used between the client and each server so that the TCP connection-setup three-way handshake delay is avoided and also so that several slow-start phases are avoided. This paraloading scheme, as opposed to the history-based and semi-dynamic parallel access paraloading methods examined before, is not based on the HTTP protocol (using the HTTP byte range header to download a block of a file), but uses a proprietary paraloading server and client running on top of the TCP transport layer protocol. The file size is obtained from the first server with which the client comes into contact with and thus no time is lost by performing separate server probing for the explicit purpose of obtaining the file size value. The file, as mentioned above, is divided into fixed-size blocks and a request is made to each server to download a distinct block.

3 Once a given server has completed the assigned file block download a new block for downloading is assigned to it. Currently the block size used is set to 1 Mbyte. The selection of the block size is an important issue, in which three facts should be taken into account: The file block size should be such, such that the number of blocks is larger than the number of servers. Otherwise the faster servers will exhibit large idle times. Each block should be small enough so as to provide a fine enough granularity, for the same reason as mentioned in i). On the other hand, each block should be large enough so as to reduce the number of requests that need to be made to servers, thus reducing the ratio of idle time to download time. The major point of difference between this paraloading algorithm and semi-dynamic paraloading, is that the number of mirror servers used in downloading does not remain static. After a connection is established to each server and a block download request is made to every server, the so-called server downscaling testing commences in the case that 4 or more servers are currently being used. With downscaling testing, the transmission rate of every server is monitored during the parallel download. At given time intervals, the slowest of the servers (based on the recorded transmission rates) is selected to remain idle for a period of time by not being given any new block download requests. After the testing time has elapsed, the aggregate download transmission rate (i.e. the sum of the individual transmission rates of each server) is compared to the aggregate data rate of the time before the selected server was made idle. If the new aggregate rate is lower than the old aggregate rate by less than a certain threshold percentage (currently a value of 15% is used, but can be varied if desired) or the new rate happens to be equal or even higher, then the server is deemed to be unnecessary since it offers no substantial increase in bandwidth and is deleted from the list of active mirror servers for the current download and is used no further in downloading, thus freeing up unnecessarily tied up server and network resources. In this case server downscaling testing will continue by proceeding to the next server (which will be the slowest among the currently active servers, as before). Otherwise, if the drop in the aggregate download rate is above the threshold percentage, then the server is taken out of the idle state and used again in downloading. In addition, server downscaling will cease in this case since it is considered that we have now reached the ideal number of servers needed i.e. the least number of servers required to give the maximum possible download rate. Server downscaling will also terminate at any time if less than four servers are currently active participating in the download process. After server downscaling has terminated (i.e. no more servers will be deactivated in the particular download), server upscaling testing will commence. Namely, if a significant decrease in the aggregate download rate exists for a sustained period of time then an additional server will be added for use in paraloading. More specifically, at periodic time intervals (currently every 3 seconds, although this value can be adjusted) testing will start by monitoring the aggregate download rate twice at given intervals (1 seconds after test commencement and 1 seconds after that, again these values can be varied). If at both of those time intervals there is a significant drop in the aggregate download rate (of 15% or more, although as previous testing parameters this can be varied) then it is concluded that there is a sustained drop in the download rate and to remedy this a mirror server, if an available one exists of course, is added for use in downloading to try to increase the rate. The fact that paraloading starts using all available mirror servers and then downscales as opposed to for example starting from one server and then adding mirror servers until there was no substantial increase in bandwidth is deliberate. The reason is that the purpose is to decrease download times and thus bandwidth underutilization (using too few servers) is a much more significant factor than bandwidth overutilization (using too many servers). While bandwidth underutilization results in lost time, bandwidth overutilization simply results in temporary use of unneeded resources that will be released eventually. One observation that should briefly be made here is the duration of the testing intervals used in server downscaling testing as examined above. On the one hand it is desirable that testing periods be as short as possible so as to complete mirror server deletion as quickly as possible. On the other hand though these testing intervals must be long enough so as to obtain valid data (transmission rate readings) that can be used to make a valid decision. In other words when a server is made idle in order to measure the effects on the aggregate download rate, sufficient time must elapse to allow the network to enter into a steady-state so to speak so as to measure the real effects on the aggregate download rate. The value used currently, and believed to satisfy both requirements stated above, is 1 seconds. 3. Experimental Results and Analysis Experiments were conducted by measuring the download times using various downloading methods. More specifically, the download times using dynamic paraloading were compared to those of single server FTP file downloading and single server multiple parallel connection downloading using the dynamic paraloading client and server. In all tests conducted clients from the same domain were used along with a multitude of geographically dispersed servers in Canada and the United States. More specifically, a total of eight servers were used: three at the University of Victoria in British Colombia, two TRLabs servers in Winnipeg and Regina, two servers at the Purdue University and an additional server at the University of Illinois.

4 16 25 loading time/ (sec) hour of day (2:3 AM - 2:3 PM) Multiserver Paraloading Fastest FTP Slowest FTP loading time/ (sec) hour of day (2:3AM - 2:3PM) 19 Multiserver Paraloading Fastest Single Server Paraloading Slowest Single Server Paraloading Figure 1. Download times of multi-server dynamic paraloading and single server FTP. Figure 2. Download times of multi-server dynamic paraloading and single-server multiple connection paraloading. The tests (dynamic paraloading, single-server multiple connection paraloading and simple single-server FTP downloading) were conducted over a 24-hour period with results being obtained every hour in order to get results for the performance of each of the three downloading schemes used under relatively heavy and varying traffic conditions (during the day) and under lighter and less varying traffic conditions i.e during the early and late hours. In testing dynamic parallel access downloading, all eight remote hosts were used as mirror servers to download a 45 Mbyte file. The same 45 Mbyte file was used for testing the other two single-server downloading schemes also, testing with many different servers. For the single-server multiple connection paraloading scheme, a single mirror server was used each time with eight connections between client and server being used, i.e. as many as the number of mirror servers used in multi-server dynamic paraloading. Figure 1 compares the results obtained for multiserver dynamic paraloading with those obtained for singleserver FTP downloading (fastest and slowest cases). From the results it can be seen that while paraloading significantly outperformed the slowest FTP downloading case, being approximately 1 times faster, the difference is significantly smaller when compared to the fastest FTP download, but still much faster proving the obvious benefit of paraloading. The maximum theoretical performance of multi-server dynamic paraloading is if the aggregate download rate is equal to the sum of the individual server download rates which is not the case here. This is due to the fact that given a sufficiently large number of servers a saturation point is reached beyond which the aggregate rate can increase no more no matter how many servers are added due to a bottleneck at the receiving client and/or along the network path between client and mirror servers. This is also the reason for which server downscaling, as explained before, was added to the dynamic paraloading scheme. Figure 2 compares single-server multiple connection paraloading results to those obtained for multi-server dynamic paraloading. Here we notice that mutli-server paraloading is somewhat slower than the fastest case of singleserver multi-connection paraloading. This is not unexpected as its only natural that in multi-server paraloading the slower servers will degrade the performance and will very likely be somewhat slower than when using the same number of connections to the fastest server(s). The advantage of multi-server paraloading even in this case is the better load balancing it achieves by spreading the load across multiple servers and not just one server where with multiple connections to one server the server will very soon become congested. The case is the opposite though for the slowest single-server multi-connection paraloading, where multiserver paraloading significantly outperforms it due mainly to one of the advantages of using multiple servers which is higher performance mirror servers compensating for the lower performance servers. Figure 3 summarizes the results displayed in the previous two figures comparing multi-server dynamic paraloading performance to the best FTP and single-server paraloading performances. It is worthy of mention here that the download performances, particularly that of FTP, degrade during the morning and afternoon hours and the variation also of the download times increases during those hours. This applies mostly to FTP, with the two paraloading schemes being affected to a lesser extent, thus concluding that the use of multiple connections also has a smoothing effect on download performance isolating to a certain extent the overall download performance from the perfor-

5 loading time/ (sec) hour of day (2:3AM - 2:3PM) 19 Multiserver Paraloading Fastest Single Server Paraloading Fastest FTP by the authors of this paper was presented and examined. Dynamic parallel access downloading, based on the experiments performed, proved that it performed very well in terms of reducing download time even under varying traffic/network conditions. Additionally, dynamic parallel access possesses some advantages over the other two downloading methods briefly examined, such as being able to adjust the number of mirror servers that are active thus releasing server and network resources that are unnecessarily utilized, and also improved server load balancing compared to the other downloading schemes. It is strongly believed that performance can be further improved with additional enhancements. Some possible enhancements to dynamic parallel access paraloading that should be examined as future work are: Figure 3. Download times of multi-server dynamic paraloading and best single-server multi-connection paraloading and FTP performances. mance of any individual server or other traffic/network condition variations. 3.1 Comparison to Semi-Dynamic Paraloading From the experimental results, presented and analyzed in the previous section, it is apparent that the dynamic paraloading scheme performs very well increasing download performance significantly when compared to slower servers. One aspect were dynamic paraloading has an advantage over semi-dynamic, is the adjustable number of mirror servers that are used at any given time. Instead of using all the mirror servers that are available, dynamic paraloading can reduce the number of active servers thus releasing server and network resources that are unnecessarily utilized. Another advantage of dynamic paraloading is the better server load balancing achieved compared to the other downloading schemes. While all downloading schemes achieve server load balancing to a certain extent, by distributing the connections among all mirror servers, dynamic paraloading with its server downscaling feature releases servers that are unnecessarily utilized (i.e. that offer very little to the aggregate download rate) and this would include heavily loaded servers. The ability to be able to add additional connections between the client and a given server. More specifically, the ability to add a 2nd, 3rd etc. connection between the client and the fastest servers rather than just being able to add an additional mirror server. Use pipelining of block download requests in order to minimize the number of idle periods between block downloads. The development of a method to dynamically retrieve the list of mirror servers, such as a directory service for example. Determine the effects of paraloading in terms of network congestion. References [1] J. Byers, M. Luby,and M. Mitzenmacher, Accessing multiple mirror sites in parallel: Using tornado codes to speed up downloads, IEEE INFOCOM, [2] J. Kangasharju, K.W. Ross, and J. W. Roberts, Locating copies of objects using the domain name system, 4th International Caching Workshop, Mar. 2. [3] A. Miu and E. Shih, Performance Analysis of a Dynamic Parallel Downloading Scheme from Mirror Sites Throughout the Internet, Technical Report, Laboratory of Computer Science, MIT, 2. [4] P. Rodriguez, A. Kirpal, and E. Biersack, Parallelaccess for mirror sites in the Internet, IEEE INFO- COM, Conclusions and Future Work In this paper a new parallel access downloading scheme referred to as Dynamic Parallel Access scheme developed