Network Processors as Building Blocks in Overlay NetworksAda Gavrilovska, Karsten Schwan, Ola Nordstrom, Hailemelekot SeifuCenter for Experimental Research in Computer Systems (CERCS)Georgia Institute of TechnologyAtlanta, Georgia, 30332ada, schwan, nalo, seif@cc.gatech.eduAbstractThis paper proposes an architecture that permits se-lected application- and middleware-level functionality to be‘pushed into’ network processors. Such functionality is rep-resented as stream handlers that run on the network proces-sors (NPs) attached to the host nodes participating in over-lay networks. When using stream handlers, application-and middleware-level functionality is ‘split’ into multiplecomponents that are jointly executed by the host and theattached NP (ANP). Resulting improvements in applicationperformance are due to the network-near nature of ANPsand due to the ability to dynamically customize stream han-dlers to meet current application needs or match currentnetwork resources. The evaluation of our current prototypeimplementation indicates that the use of ANP-level handlerscan reduce the delays on the application data path by morethan 25%, and can sustain higher throughput for the ap-plication services provided by stream handlers. In addi-tion, stream handlers are a suitable basis for scalable im-plementations of data-increasing services like destination-customized multicast.1. Attached Network Processors in OverlayNetworksResearch in active networking has long argued the ben-efits of dynamically reconfiguring the network infrastruc-ture, to better match application needs with platform con-ditions [11, 16, 15]. Technical advances due to such workinclude the development of architectures and runtime meth-ods for programmable routers and switches, the provisionof general techniques for adapting communications, and ex-perimentation with applications like video distribution andservice proxies. While building on such advances, ourresearch focuses on the network’s ‘edges’, where we areinvestigating the benefits attained for applications by dy-namically customizing the functionality of ‘attached net-work processors’ (ANPs). ANPs are the programmablenetwork devices that are attached to the end hosts usedby applications. The specific ANPs used in our researchare the emerging class of programmable network proces-sors, which are becoming an attractive target for deploy-ing new functionality ‘into’ the network, at lower costsand with shorter development times than custom-designedASICs. Their utility has already been demonstrated for pro-gramming network-centric services like software routing,network monitoring, intrusion detection, load-balancing,service differentiation, and others. In contrast, we focuson application-level functionality. For example, data fil-tering and packet scheduling benefits applications like re-mote sensing, remote visualization, online collaboration,and multimedia [17, 7]. Dynamic synchronization methodsor content-based multicast benefit high performance simu-lations or transactional applications like operational infor-mation systems [5, 10]. In all such cases, application-levelperformance is directly affected by the ability to ‘push’suitable functionality onto ANPs, and to dynamically cus-tomize it to meet current application needs or match currentnetwork resources.Our past work has focused on using ANPs in systemarea networks with high end cluster machines [10]. Ourcurrent work is broader, in that it targets any distributedapplication that uses overlay networks. These include theonline collaboration and distributed scientific codes con-sidered in our earlier work, transactional applications, andthe middleware-based grid services that are becoming in-creasingly important for future distributed systems. Thespecific middleware considered in our research implementspublish/subscribe event-based communications across co-operating peers [3, 14].The key idea presented in this paper is to permit applica-tions and middleware to run overlay services on the ‘edge-deployed’ NPs that are attached to participating hosts. Thisis achieved by ‘splitting’ the middleware- or application-level functionality executed on hosts (in their OS ker-nels or at application-level) into multiple components thatProceedings of the 11 th Symposium on High Performance Interconnects (HOTI’03) 0-7695-2012-X/03 $17.00 © 2003 IEEEare efficiently and jointly executed by the host and ANP.This is shown in Figure 2, where this additional function-ality is placed into the region labeled Access,whichissplit across ANP and host. Splitting should be dynamic,(i.e., when overlays are first constructed), and split compo-nents should be runtime-configurable, so as to continuouslymatch their operation to current application needs and net-work resources (e.g., by configuring certain execution pa-rameters, such as rate or point of invocation, or by configur-ing application-level parameters, such as filtering thresholdsor bounding box coordinates).The goal is to use ANP resources to deliver to a widerange of scientific and commercial applications improve-ments in performance or reliability. Benefits experiencedby applications are due to customized communications andthe reduced computational loads on end-host, e.g., by of-floading overlay computations onto ANPs. We avoid someof the safety and security issues raised in the context of ac-tive networks by engaging end host operating systems in thecontrol of ANP-level functionality.The experimental results presented in this paper use In-tel’s’ IXP1200 network processors as ANPs. Measurementsshow that use of ANPs vs. hosts for certain application-level services can reduce the latency of the application datapath by more than 25%, and that it can sustain high through-put for most of the services considered. It can also provide a25% improvement for certain data-increasing services, suchas destination-customized multicast.Remainder of paper. In the remainder of paper, we firstdescribe the application domains and platforms targeted byour work. Section 3 briefly describes the ANP-resident por-tion of an application-specific service, termed stream han-dlers, then present the system components of the compositehost-ANP nodes. Section 4 presents and discusses experi-mental results attained with our prototype implementation.Related work, conclusions, and future work are presentedlast.2. Application Domains and Execution
View Full Document
Unlocking...