`
`The MAGIC Project: From Vision to Reality
`
`BarbaraFuller, Mitretek Systems
`Ira Richer, Corporation for National Research Initiatives
`
`Abstract
`In the MAGIC project, three major components — an ATM internetwork, a
`distributed, network-based storage system, and a terrain visualization applica-
`tion — were designed, implemented, and integrated to create a testbed for
`demonstrating real-time, interactive exchange of data at high speeds among
`distributed resources. The testbed was developed as a system, with special
`consideration to how performance was affected by interactions among the
`components. This article presents an overview of the project, with emphasis
`on the challenges associated with implementing a complex distributed system,
`and with coordinating a multi-organization collaborative project that relied on
`distributed development. System-level design issues and performance measure-
`ments are described, as is a tool that was developed for analyzing perfor-
`mance and diagnosing problemsin a distributed system. The management
`challenges that were encountered and someof the lessons learned during the
`course of the three-year project are discussed, and a brief summary of
`MAGIC.Il, a recently initiated follow-on project,
`is given.
`
`to-end system performancerather than individual compo-
`igabit-per-second networks offer the promise
`nent performance.
`of a major advance in computing and commu-
`The objective of the MAGIC (which stands for “Multidi-
`nications: high-speed access to remote
`mensional Applications and Gigabit Internetwork Consor-
`resources, including archives, time-critical
`tium”) project was to build a testbed that could
`data sources, and processing power. Over the past six
`demonstrate real-time, interactive exchange of data at
`years, there have beenseveral efforts to develop gigabit
`gigabit-per-second rates among multiple distributed
`networks and to demonstratetheir utility, the most notable
`resources. This objective was pursued through a multidisci-
`being the five testbeds that were supported by ARPA and
`National Science Foundation (NSF) funding: Aurora,
`plinary effort involving concurrent development and subse-
`quent integration of three testbed components:
`BLANCA, CASA, Nectar, and VISTAnet [1]. Each of
`¢ An innovative terrain visualization application that
`these testbeds comprised a mix of applications and net-
`requires massive amounts of remotely stored data
`working technology, with some focusing more heavily on
`applications and others on networking. The groundbreak-
`¢ A distributed image server system with performance suf-
`
`ing work donein these testbeds hadasignificant impact ficient to support the terrain visualization application
`on the development of high-speed networking technology
`¢ A standards-based high-speed internetworkto link the
`and onthe rapid progress in this area in the 1990s.
`computing resources required for real-time rendering of
`the terrain
`It became clear, however, that a new paradigm for
`application development was neededin order to realize
`The three-year project began in mid-1992 and involved
`the full benefits of gigabit networks. Specifically, network-
`the participation, support, and close cooperation of many
`based applications and their supporting resources. such as
`diverse organizations from government, industry, and
`data servers, must be designed explicitly to operate effec-
`academia. These organizations had complementaryskills
`tively in a high-speed networking environment. For exam-
`andhadthe foresight to recognize the benefits of collabo-
`ple, an interactive application working with remote storage
`ration. The principal MAGICresearchparticipants were:
`devices must compensate for network delays. The MAGIC
`¢ Earth Resources Observation System Data Center, U.S.
`Geological Survey (EDC)!
`project, which is the subject of this article, is the first high-
`speed networking testbed that was implemented according
`¢ Lawrence Berkeley National Laboratory, U.S. Depart-
`ment of Energy (LBNL)!
`to this paradigm. The major componentsofthe testbed
`* Minnesota Supercomputer Center, Inc. (MSCI)!
`were considered to be interdependentparts of a system,
`¢ MITRE Corporation!
`and whereverpossible they were designed to optimize end-
`¢ Sprint
`¢ SRI International (SRI)!
`
`The work reportedhere wasperformed while the authors were with the
`MITRE Corp. in Bedford, MA, and was supported by the Advanced
`Research ProjectAgency (ARPA) under contract F19628-94-D-001.
`
`! These organizations were fundedby ARPA.
`
`TEBE NeBae MalPues£448
`
`0890-8044/96/$05.00 © RIAHFFEd Patents Exhibit 1005 App'x A-N®
`
`
`
`
`
`Image server system
`(Storage and transmission
`of raw image tiles)
`
`Distributed processing
`(Real-time image processing)
`
`—
`[Bic
`
`Image server system
`(Storage and transmission
`of processed tiles)
`
`Rendering engine
`(Rendering and visualization
`of terrain)
`
`Workstations
`(Over-the-shoulder
`
`
`
`™ Figure 1. Planned functionality ofthe MAGICtestbed.
`
`Overview of the MAGIC Testbed
`
`Or of the primary goals of the MAGICproject was to
`create a testbed to demonstrate advanced capabilities
`that would not be possible without a very high-speed inter-
`network. MAGIC accomplished this goal by implementing
`an interactive terrain visualization application, TerraVi-
`sion, that relies on a distributed image server system (ISS)
`to provide it with massive amountsof data in real time.
`The planned functionality of the MAGICtestbed is depict-
`ed in Fig. 1. Currently, TerraVision uses data processed
`off-line and stored on the ISS. In the future the applica-
`tion will be redesigned to enable real-time image process-
`ing as well as real-time terrain visualization (see the last
`section). Note that the workstations which house the appli-
`cation, the servers of the ISS, and the “over-the-shoulder”
`tool (see subsection entitled “The Terrain Visualization
`Application”), as well as those that will perform the on-
`line image processing, can reside anywhere onthe network.
`
`The MAGIC Internetwork
`The MAGICinternetwork, depicted in Fig. 2, includes six
`high-speed local area networks (LANs) interconnected by
`a wide area network (WAN) backbone. The backbone,
`which spansa distance of approximately 600 miles, is
`based on synchronous optical network (SONET)technolo-
`gy and provides OC-48 (2.4 Gb/s) trunks, and OC-3 (155
`Mb/s) and OC-12 (622 Mb/s) access ports. The LANs are
`based on asynchronous transfer mode (ATM)technology.
`Five of the LANs — those at BCBL in Fort Leavenworth,
`Kansas, EDC in Sioux Falls, South Dakota, MSCI in Min-
`neapolis, Minnesota, Sprint in Overland Park, Kansas, and
`U S WESTin Minneapolis, Minnesota — use FORE Sys-
`tems models ASX-100 and ASX-200 switches with OC-3c
`and 100 Mb/s TAXIinterfaces. The ATM LANat KUin
`Lawrence, Kansas, uses a DEC AN2switch, a precursor to
`the DEC GigaSwitch/ATM, with OC-3c interfaces. The
`network uses permanentvirtual circuits (PVCs) as well as
`switched virtual circuits (SVCs) based on both SPANS, a
`FORESystemssignaling protocol, and the ATM Forum
`User-NetworkInterface (UNI) 3.0 Q.2931 signaling stan-
`
`view of terrain)
`
`
`
`* University of Kansas (KU)!
`¢ US WEST Communications, Inc.
`Other MAGICparticipants that contributed equipment,
`facilities, and/or personnel to the effort were:
`¢ Army High-Performance Computing Research Center
`(AHPCRC)
`¢ Battle Command Battle Laboratory, U.S. Army Com-
`bined Arms Command (BCBL)
`* Digital Equipment Corporation (DEC)
`¢ Nortel, Inc./Bell Northern Research
`¢ Southwestern Bell Telephone
`¢ Splitrock Telecom
`This article presents an overview of the
`MAGICproject with emphasis on the chal-
`lenges associated with implementing a
`complex distributed system. Companion
`articles [2, 3] focus on a LAN/WANgate-
`way and a performanceanalysis tool that
`were developed for the MAGIC testbed.
`Thearticle is organized as follows. Thefol-
`lowing section briefly describes the three
`major testbed components: the internet-
`work, the image server system, and the
`application. The third section discusses
`some of the system-level considerations
`that were addressed in designing these
`components,and the fourth section pre-
`sents some high-level performance mea-
`surements. The fifth (affectionately
`entitled “Herding Cats”) and sixth sections
`describe howthis multi-organizational col-
`laborative project was coordinated, and the
`technical and managerial lessons learned.
`Finally, the last section provides a brief
`summary of MAGIC-II, a follow-on project
`begunin early 1996.
`
`— SONET OC-48
`— SONET OC-12 or OC-3
`Workstations include:
`DEC, SGI, SUN for ISS and
`over-the-shoulder, SGI for
`terrain visualization
`
`
`
`16 Unified Patents ExhibRE?00%AppxuteNp6Page 148 of 448
`
`
`
`™@ Figure 2. Configuration ofthe MAGIC ATM internetwork.
`
`
`
`
`
`
`
`
`Image tiles of
`terrain data
`
`
`@ Figure 3. Relationship betweentile resolutions andperspective view.
`(Source: SRIInternational)
`
`dard. The workstations at the MAGIC sites
`include models from DEC, SGI, and Sun. As
`part of MAGIC, an AN2/SONETgateway with
`an OC-12c interface was developed to link the
`AN2 LAN at KU to the MAGIC backbone[2].
`In addition to implementing the internetwork,
`a variety of advanced networking technologies
`were developed and studied under MAGIC. A
`high-performance
`parallel
`interface
`(HIPPI)/ATM gateway was developedtointer-
`face an existing HIPPI network at MSCIto the
`MAGIC backbone. The gateway is an IP router rather
`than a network-layer device such as a broadbandintegrat-
`ed services digital network (B-ISDN) terminal adapter,
`and was implementedin software on a high-performance
`workstation (an SGI Challenge). This architecture provides
`Data Preparation — In orderto render an image, TerraVi-
`a programmable platform that can be modified for net-
`sion requires a digital description of the shape and appear-
`workresearch, and in the future can readily take advan-
`ance of the subject terrain. The shape oftheterrain is
`tage of more powerful workstation hardware. In addition,
`represented by a two-dimensional grid of elevation values
`the platform is general-purpose; that is, it is capable of
`knownas a digital elevation model (DEM). The appearance
`supporting multiple HIPPI interfaces as well as other
`interfaces such as fiber distributed data interface (FDDI).
`of the terrain is represented byaset of aerial images,
`knownas orthographic projection images (ortho-images),
`Software was developed to enable UNIX hosts to com-
`municate using Internet Protocol (IP) over an ATM net-
`that have been specially processed(i.e., ortho-rectified) to
`work. This
`IP/ATM software currently runs on
`eliminate the effects of perspective distortion, and are in
`SPARCstations under Sun OS4.1 and includes a device
`precise alignment with the DEM. Tofacilitate processing,
`driver for the FORE SBAseries of ATM adapters.It sup-
`distributed storage, and high-speed retrieval over a net-
`ports PVCs, SPANS, and UNI3.0 signaling, as well as the
`work, the DEM andimagesare divided into small fixed-
`size units knownasfiles.
`“classical” IP and Address Resolution Protocol (ARP)
`Low-resolution tiles are required for terrain thatis dis-
`over ATM model[4]. The software should be extensible to
`other UNIX operating systems, ATM interfaces, and
`tant from the viewpoint, whereas high-resolution tiles are
`IP/ATM address-resolution and routing strategies, and will
`required for close-in terrain. In addition, multiple resolu-
`facilitate research on issues associated with the integration
`tions are required to achieve perspective. These require-
`of ATM networks into IP internets.
`ments are addressed by preparing a hierarchy of
`In order to enhance network throughput, flow-control
`increasingly lower-resolution representations of the DEM
`schemes were evaluated and applied, and IP/ATM host
`and ortho-imagetiles in which each level is at half the res-
`parameters were tuned. Experiments showed that through-
`olution of the previous level. The tiled, multiresolution
`put close to the maximum theoretically possible could be
`hierarchy and the use of multiple resolutions to achieve
`attained on OC-3 links over long distances. To achieve
`perspective are shownin Fig.3.
`high throughput, both the maximum transmission unit
`Rendering of the terrain on the screen is accomplished
`(MTU)and the Transmission Control Protocol (TCP) win-
`by combining the DEM and ortho-image tiles for the
`dow mustbe large, and flow control must be used to
`selected area at the appropriate resolution. As the user
`ensure fairness and to avoidcell loss if there are interact-
`travels over the terrain, the DEMtiles and their corre-
`ing traffic patterns [5, 6].
`sponding ortho-imagetiles are projected onto the screen
`using a perspective transform whose parametersare deter-
`mined by factors such as the user’s viewpoint andfield of
`view. The mapping of a transformed ortho-imagetoits
`DEMand the rendering of that image are shownin Fig.4.
`The data set currently used in MAGIC covers a 1200
`km? exercise area of the National Training Center at Fort
`Irwin, California, and is about 1 Gpixelin size. It is derived
`from aerial photographs obtained from the National Aerial
`Photography Program archives and DEM data obtained
`from the U.S. Geological Survey. The images are at
`approximately 1 m resolution (i.e., the spacing between
`pixels in the image corresponds to 1 m on the ground).
`The DEM dataare at approximately 30 m resolution(i.e.,
`elevation values in meters are at 30 m intervals).
`Software for producing the ortho-images and creating
`the multiresolution hierarchy of DEM and ortho-image
`tiles was developed as part of the MAGIC effort. These
`processes were performed “off-line” on a Thinking
`Machines Corporation Connection Machine (CM-5) super-
`computer owned by the AHPCRCandlocated at MSCI.
`The tiles were then stored on the distributed servers of the
`ISS and usedby terrain visualization software residing on
`rendering engines at several locations.
`
`cation, and rendering the image. MAGIC’s approach to
`accomplishing these activities is described below. Enhance-
`ments to the application that provide additional features
`and capabilities are also described.
`
`The Terrain Visualization Application
`TerraVision allows a user to view and navigate through(i.e.,
`“fly over”) a representation of a landscape created from
`aerial or satellite imagery [7]. The data used by TerraVision
`are derived from raw imagery and elevation information
`which have been preprocessed by a companion application
`known as TerraForm. TerraVision requires very large
`amounts of data in real time, transferred at both very
`bursty and high steady rates. Steady traffic occurs when a user
`moves smoothly through the terrain, whereas bursty traffic
`occurs when the user jumps (“teleports”) to a new position.
`TerraVision is designed to use imagery data that are locat-
`ed remotely and supplied to the application as needed by
`meansof a high-speed network. This design enables Ter-
`raVision to provide high-quality, interactive visualization
`of very large data sets in real time. TerraVision is of direct
`interest to a variety of organizations, including the Depart-
`ment of Defense. For example, the ability of a military
`officer to see a battlefield and to share a common view
`with others can be very effective for commandandcontrol.
`Terrain visualization with TerraVision involves two activi-
`ties: generating the digital data set required by the appli-
`
`TEEE NeBradse Mafiigsf'94a48
`
`Unified Patents Exhibit 1005 App'x A-N’
`
`
`
`
`ell
`
`Aerial terrain image
`
`Elevation model
`
`possible by precisely aligning the DEM and imagery
`data with a world coordinate system as well as with
`each other.
`A numberof buildings and vehicles have been
`created and stored on the rendering enginefordis-
`play as an overlay on theterrain. The locations of
`vehicles can be updatedperiodically by transferring
`vehicle location data, acquired with a global posi-
`tioning system receiver, to the rendering engine for
`integration into the terrain visualization displays.
`=
`eeae
`1?
`~
`<
`s=
`~_-
`Registration of the user’s viewpoint to a map enables
`
`
`— = boaBd= ba onthe .. ae
`
`
`
`>
`tl
`ee
`tT
`>
`the user to specify the area he wishes to explore by
`—w 4 Ls
`_—
`-
`=
`~~.
`pointing to it, and it aids the user in orienting him-
`self.
`In addition, an over-the-shoulder (OTS) tool was
`developedto allow a user at a remote workstation to
`view the terrain as it is rendered. The OTStoolis
`based on a client/server design and uses XWindow
`system calls. The user can view the entire image on
`the SGI screen at low resolution, and can also select
`a portion of the screen to view at higher resolution.
`The framerate varies with the size and resolution of
`the viewed image, and with the throughput of the
`workstation.
`
` ;
`= _
`
`Aerial terrain image is mapped onto elevation model
`
`-_—
`
`ee ae
`
`The Image Server System
`Elevation data rendered with orthographic imageofterrain
`The ISS stores, organizes, and retrieves the pro-
`cessed imagery and elevation data required by Ter-
`®@ Figure 4. Mapping an ortho-image ontoits digital elevation model.
`raVision for interactive rendering of the terrain. The
`(Source: SRIInternational)
`ISS consists of multiple coordinated workstation-
`based data servers that operate in parallel and are
`designed to be distributed around a WAN. This
`architecture compensatesfor the performancelimitations
`of current disk technology. A single disk can deliver data
`at a rate that is about an order of magnitude slower than
`that needed to support a high-performance application
`such as TerraVision. By using multiple workstations with
`multiple disks and a high-speed network, the ISS can deliv-
`er data at an aggregate rate sufficient to enable real-time
`rendering of the terrain. In addition, this architecture per-
`mits location-independentaccess to databases, allows for
`system scalability, and is low in cost. Although redundant
`arrays of inexpensive disks (RAID) systems can deliver higher
`throughput thantraditional disks, unlike the ISS they are
`implemented in hardware and, as such, do not support multi-
`ple data layout strategies; furthermore, they are relatively
`expensive. Such systemsare therefore not appropriate for
`distributed environments with numerousdata repositories
`serving a variety of applications.
`The ISS, as currently used in MAGIC, comprises four or
`five UNIX workstations (including Sun SPARCstations,
`DEC Alphas, and SGI Indigos), each with fourto six fast
`SCSI disks on two to three SCSI host adapters. Each serv-
`er is also equipped with either a SONET or a TAXInet-
`workinterface. The servers, operating in parallel, access
`the tiles and send them over the network, which delivers
`the aggregate stream to the host. This processis illustrated
`in Fig. 5. More details about the design and operation of
`the ISS can be foundin [8].
`
`Image Rendering — TerraVision provides for two modes
`of visualization: two-dimensional (2-D) and three-dimen-
`sional (3-D). The 2-D modeallowsthe userto fly over the
`terrain, looking only straight down. The user controls the
`view by means of a 2-D input device such as a mouse.
`Since virtually no processing is required, the speed at
`which imagesare generated is limited by the throughput of
`the system comprising the ISS, the network, and the ren-
`dering engine.
`In the 3-D mode, the usercontrols the visualization by
`meansof an input device that allows six degreesof free-
`dom in movement. The 3-D mode is computationally
`intensive, and satisfactory visualization requires both high
`framerates(i.e., 15-30 frames/s) and lowlatencies(i.e., no
`more than 0.1 s between the time the user moves an input
`device and the time the new frame appears onthe screen).
`High framerates are achieved by using a local very-
`high-speed rendering engine, an SGI Onyx, with a cache of
`tiles covering not only the area currently visible to the
`user, but also adjacent areas that are likely to be visible in
`the near future. A high-speed search algorithm is used to
`identify the tiles required to render a given view. For
`example, as noted above, perspective (i.e., 3-D) views
`require higher-resolution tiles in the foreground and
`lower-resolution tiles in the background. TerraVision
`requests the tiles from the ISS, places them in memory, and
`renders the view. Latency is minimized by separating image
`rendering from data input/output (I/O) so that the twoactiv-
`ities can proceed simultaneously rather than sequentially
`(see the section entitled “Design Considerations”).
`
`Additional Features and Capabilities — TerraVision
`includes two additional features: superposition of fixed
`and mobile objects on the terrain, and registration of the
`user’s viewpoint to a map. Both of these features are made
`
`Design Considerations
`MAGIC, the single most perspicuouscriterion of suc-
`cessful operation is that the end user observessatisfactory
`performanceofthe interactive TerraVision application.
`Whenthe userflies over the terrain, the displayed scene
`must flow smoothly, and when heteleports to an entirely
`
`
`
`18 Unified Patents ExhibRE?00%AppxuteNp6Page 150 of 448
`
`
`
`
`
`Tiles intersected
`by the ee oftravel
`4
`
`
`
`transmission to application
`
`
`
`Location tiles along path on
`ISS servers/disks
`
`Server and disk
`——-» S1D1
`—» S1D2
`—— $2D1
`® —-» 5S1D1
`——¥» S2D2
`———» S1D2
`Parallel retrieval of tiles and
`——¥ $2D1
`
`Server 2
`
`os
`ceepeee
`high-sarouution
`
`different location, the new scene must
`appear promptly. Obtaining such per-
`formance mightberelatively straight-
`forward if the terrain data were
`collocated with the rendering engine.
`However,one of the original premises
`underlying the MAGICprojectis that
`the data set and the application are
`not collocated. Thereare several rea-
`sons for this, the most important being
`that the data set could be extremely
`large, so it might not be feasible to
`transfer it to the user’s site. Moreover,
`experience has shownthat in many
`cases the “owner” of a datasetis also
`its “curator” and maybe reluctant to
`distribute it, preferring instead to keep
`the data locally to simplify mainte-
`nance and updates. Finally, it was
`anticipated that future versions of the
`application might work with a mobile
`user and with fused data from multiple
`sources, and neither of these capabili-
`ties would be practical with local data.
`Therefore, since the data will not be
`local, the MAGIC components must
`be designed to compensate for possible
`delays and other degradations in the
`end-to-end operation of the system.
`In order to understand system-level
`design issues, it is necessary to outline
`the sequence of events that occurs when the user moves the
`input device, causing a new sceneto be generated. TerraVi-
`sion first producesa list of new tiles required for the scene.
`This list is sent to an ISS master, which performs a name
`translation, mapping the logical address ofeachtile (the
`tile identifier) to its physical address (server/disk/location
`on disk). The master then sends each server an orderedlist
`of the tiles it must retrieve. The server discards the previ-
`ouslist (even if it has not retrieved all the tiles on that
`list) and begins retrieving the tiles on the new list. Thus,
`the design for the system comprising TerraVision, the ISS,
`and the internetwork must address the following questions:
`¢ How can TerraVision compensate fortiles it needs for
`the next image but have not yet been received?
`¢ Howoften should TerraVision request tiles from the ISS?
`Where should the ISS master be located?
`How should tiles be distributed among the ISS disks?
`«How cancell loss be minimized near the rendering site
`wherethetile traffic becomes aggregated and conges-
`tion may occur?
`
`™@ Figure 5. Schematic representation ofthe operation ofthe ISS. (Source: Lawrence
`Berkeley National Laboratory)
`
`used in place of the 16 level-3 tiles. This substitution mani-
`fests itself by the affected portion of the rendered image
`appearing “fuzzy” for a brief period of time. Temporary
`substitution of low-resolution tiles for high-resolution tiles
`is particularly effective for teleporting because that opera-
`tion requires a large number of newtiles, so it is more
`likely that one or morewill be delayed.
`Second, TerraVision attemptsto predict the path the
`user will follow, requesting tiles that might soon be need-
`ed, and assigning one of three levels of priority to eachtile
`requested. Priority-1 tiles are needed as soon aspossible;
`the ISS retrieves and dispatches thesefirst. This set of tiles
`is ordered by TerraVision, with the coarsest assigned the
`highest priority within the set. The reasonsare:
`¢ The rendering algorithm needsthecoarsetiles beforeit
`needs the next-higher-resolutiontiles.
`¢ There are fewertiles at the coarser resolutions, so it is
`less likely that they will be delayed.
`Thepriority-2 tiles are those that the ISS should retrieve
`but should transmit only if there are no priority-1 tiles to
`be transmitted; that is, priority-2 tiles are put on a lower-
`priority transmit queue in the I/O buffer of each ISS serv-
`er. (ATM switches would be allowed to dropthecells
`carrying these tiles.) Priority-3 tiles are those that should
`be retrieved and cachedat the ISSserver; these tiles are
`less likely to be needed by TerraVision. Note that there is
`a trade-off between “overpredicting” — requesting too
`many tiles — which would result in poor ISS performance
`and high network load, and “underpredicting,” which
`would result in poor application performance.
`Finally, a tile will continue to be included in Terra-
`Vision’s request list if it is still needed and has not
`yet been delivered. Thus, tiles or tile requests that are
`dropped or otherwise “lost” in the networkwill likely be
`delivered in response to a subsequent request from the
`application.
`
`Missing Tiles
`Network congestion, an overload at an ISS server, ora
`componentfailure could result in the late arrival or loss of
`tiles that are requested by the application. Several mecha-
`nisms were implementedto deal with this problem. First,
`although the entire set of high-resolution tiles cannot be
`collocated with the application, it is certainly feasible to
`store a complete set of lower-resolution tiles. For example,
`if the entire data set comprises 1 Tbyte of high-resolution
`tiles, then all of the tiles that are five or more levels coarser
`would occupyless than 1.5 Mbyte, a readily affordable
`amountof local storage. If a tile with resolution at, say,
`level 3 is requested but not delivered in time for the image
`to be rendered, then, until the missing level-3 tile arrives,
`the locally available coarser tile from level 5 would be
`
`TEEE Neos¢ Ma&/Jugsf'94a4§
`
`Unified Patents Exhibit 1005 App'x A-N9
`
`
`
`
`
`Fora typical
`MAGIC configura-
`
`tion, the interval
`
`between requests
`
`is currently set at
`
`200 ms, a value
`
`that was found
`
`empirically to yield
`
`satisfactory
`
`performance.
`
`aggregate throughput which can be
`obtained from the ISS. The data place-
`ment strategy depends onthe application
`andis a function of data type and access
`patterns. For example, the retrieval pat-
`tern for a database of video clips would be
`quite different from that for a database of
`images. A strategy was developed fora
`terrain visualization type of application
`that minimizes the retrieval time for a set
`of tiles: the tiles assigned to a given disk are
`as far apart as possiblein the terrain in
`order to maximize parallelism by minimizing
`the probability that tiles on a requestlist
`are on the same disk; and on each disk,
`tiles that are near each otherin the terrain are
`placed as close as possible to minimize
`retrieval time. Although this was shown to be
`an optimal strategy for terrain path-fol-
`lowing as in TerraVision [9], it was subse-
`quently shownthat ISS performance with
`random placementoftiles was only slightly
`worse. This was partly becausetile retrieval
`time is much less than the latency in the
`ISS servers and network transit time, and is therefore not
`currently a significant factor in overall performance. Ran-
`dom placementis simpler to implement and is expected to
`be satisfactory for many other applications. However, as
`discussed for the location of the ISS master, this conclu-
`sion may haveto berevisited if the performance or the
`geographic distribution of system components changessig-
`nificantly.
`
`Frequency of Requests
`Anothertrade-off pertains to the frequen-
`cy at which TerraVision sends its request
`list to the ISS. If the interval between
`requests is too large, then sometiles will
`not arrive when needed, resulting ina
`poor-quality display; in addition, the ISS
`will be idle and hencenot usedefficiently.
`Onthe other hand, if the interval is too
`short, then the request list might contain
`tiles that are currently in transit from
`servers to the application; this would
`result in poor ISS performance and redun-
`dant networktraffic. For a typical MAGIC
`configuration,
`the interval between
`requests is currently set at 200 ms, a value
`that was found empirically to yield satis-
`factory performance. This value is based
`roughly on the measuredlatency of the
`ISS (about 100 ms) and on the estimated
`time required for a tile request to travel
`through the network from the TerraVision
`
`host to the ISS master and then to the
`mostdistant ISS server, plus the time for
`the tile itself to travel back to the host (perhapsa total of
`50 ms). Additional measurements and analysis are needed
`to more precisely determine the appropriate request fre-
`quencyas a function of the performanceandlocation of
`system components and of network parameters.
`
`Location of ISS Master
`Since tile requests flow from TerraVision to the ISS mas-
`ter and thence to the servers themselves, the time for
`delivering the requests to the servers is minimized when
`the masteris collocated with the TerraVision host. Howev-
`er, locating the master with the host is neither desirable
`nor practical for several reasons. The masteris logically
`part of the ISS; therefore, its location should not be con-
`strained by the application. Also, an ISS may be used with
`several applications concurrently, by multiple simultaneous
`users of a particular application, or by a user whose host
`may be unableto supportany ISS functionality (e.g., a
`mobile user). Moreover, replication of the master would
`introduce problemsassociated with maintaining consisten-
`cy among multiple masters whenthe ISSis in a read/write
`environment, as it would be whenreal-time data are being
`stored on the servers.
`To first order, the delivery time oftile requests is limit-
`ed by the time t for a request to travel from TerraVision
`to the ISS server most distant from the TerraVision host.
`Hence, if the master is approximately on the path from the
`TerraVision host to that server, then t will not be much
`greater than when the master and hostare collocated. Fur-
`thermore, in the current MAGICtestbed, t is much smaller
`than the sum of the disk latency and the networktransit
`time. In other words, there is considerable freedom in
`choosing the location of the ISS master. Satisfactory sys-
`tem performance has been demonstrated, for example,
`with the TerraVision host in Kansas City, the ISS master
`in Sioux Falls, and servers in Minneapolis and Lawrence.
`Of course, this conclusion might change if faster servers
`reduce ISS latency considerably, or the geographic span of
`the network weresubstantially larger.
`
`Distribution of Tiles on ISS Servers
`The manner in which data are distributed among the
`servers determines the degreeof parallelism and hence the
`
`Avoiding Cell Loss
`Wheninitially implemented, the MAGIC internetwork
`exhibited very low throughputin certain configurations.
`One cause of the low throughput was found to be mis-
`matches between the burst rates of components in the com-
`munications path. Examples of such rate mismatches were:
`¢« An OC-3 workstation interface transmitting cells at full
`rate across the network to a 100 Mb/s TAXIinterface
`on another workstation
`¢ Two or more OC-3 input ports at an ATM switch send-
`ing data to the same OC-3 outputport
`A mismatch, coupled with small buffers at the output
`ports of ATM switches, caused cells to be dropped, which
`in turn resulted in the retransmission of entire TCP pack-
`ets, exacerbating the problem. In some cases the measured
`useful throughput was less than one percentof the capaci-
`ty of the lower speedline.
`Previously it was noted that in many cases a large
`MTUcan increase throughput. However, once again there
`is a trade-off. As the MTUsize is increased, the number
`of ATM cells needed to carry the MTUincreases. The
`probability that one or morecells from the MTUwill be
`dropped by the networktherefore increases, which in turn
`increases the probability that the MTU will have to be
`retransmitted, thus possibly decreasing the effective
`throughput. Flow-control techniques together with large
`switch buffers and proper choice of protocol parameters
`did provide satisfactory performance. N