Integrated office systems over LANs — a performance study

Integrated office systems over LANs — a performance study

Integrated office systems over LANs-- a performance study Nicolas Georganas* and Najah Naffah t compare the performance of three LANs in relation to ...

619KB Sizes 2 Downloads 58 Views

Integrated office systems over LANs-- a performance

study Nicolas Georganas* and Najah Naffah t compare the performance of three LANs in relation to typical office system applications

In this paper, an integrated office system environment is modelled and studied. Composed of multi-media workstations, printer servers, database servers, electronic mail servers etc., the office information system (OIS) is interconnected by a I.AN. The performance of three selected networks namely Appletalk, Starlan and Ethernet, in handling the typical office system applications, is evaluated. The study applies discrete-event computer simulation, using the QNAP2 simulation software. Results comparing the performance and utilization of the three networks are presented and some useful conclusions obtained. Keywords: local area networks, CSMA, computer simulation, performance evaluation, office information systems

Many office systems today, whether created from a set of personal computers or traditional workstations (WS), are being interconnected to constitute what is now called a 'departmental' or 'group' solution. Ethernet I' 2, which in 1981 was one of the first LANs to be announced at a data transfer rate of 10 Mbit/s, has been used in several thousand sites. Many studies 2 and some simple observations have shown that it is 'underused' (1% or 0.1% usage), especially when it interconnects between 5 to 25 workstations (personal computers (PCs), word processors (WPs), etc.) exchanging files and short text messages with printing file and mail servers. Less expensive networks providing lower bandwidth have appeared, introducing a complete set of non-standard network protocols (e.g. *Department of Electrical Engineering, University of Ottawa, Ontario, Canada KIN 6N5 tAdvanced Studies Division, Bull Transac, BP 92, 91301-Massy, France

Novell, Omninet, and IBM's PC-Net based on Sytek technology at 2 Mbit/s). Two networks recently announced are Apple's Appletalk 3 at 230.4 kbit/s, and ATT's Starlan" at 1 Mbit/s. From the applications side can be observed a growing number of workstations with graphic features, and some with voice input for WP application or short message exchange (voice mail): for example, PCs such as the ATT Unix-PC, IBM PC-AT and Apple Mac, and WSs such as the Xerox Star, Sun WS, Apollo and BulI-SPS7. These workstations are connected to Non Impact Printing (NIP) servers and shared multimedia databases. With such a powerful environment, data transfer will increase substantially, potentially causing additional strain on the new generation of LANs, which are decreasing in bandwidth. In this paper, a set of present and future applications has been selected, such as the manipulation of multimedia data (i.e. storage and retrieval), a multimedia database, formatting and printing, and electronic mail. For each application, assumptions have been made about the nature of the traffic which would be generated in typical departmental configurations. Three networks are studied to see how they match the various needs of each application; the selected networks are Ethemet, Starlan and Appletalk. A typical office information system (OIS)is presented, and general characteristics of the various applications provided. The three networks are briefly outlined, and a description of QNAP2 s, the Queueing Network Analysis Program which has been applied to this study, is given. Finally, results are presented showing the performance of each network in each of the selected application environments, and some useful conclusions obtained. All the traffic studied was generated in asynchronous mode and does not include any real-time audio and video

0140-3664/87/060291-06 $03.00 © 1987 Butterworth & Co (Publishers) Ltd vol 10 no 6 december 1987

291

traffic, since the latter type of traffic causes a constant and heavy data load, which would saturate the three networks selected.

OFFICE SYSTEMS A N D THEIR TRAFFIC A typical office system (see Figure 1) comprises a set of workstations (WSs) with graphic capabilities, with or without local disc storage units. This is integral to this study, because in the latter case a WS sequentially browsing a multimedia document through the network, would generate a heavy traffic between the server and the WS duringthe session; a complete session taking 5 min to 1 h or more. In the former case, the browsing is done locally, and may necessitate occasional transfer of a copy of a document from the file server to the WS disc. In addition to the WS, there might be a scanner which digitizes the images of the various objects (e.g. pages of documents, pictures, solid objects) which are of interest to the WS users. The scanner might be hooked onto the network and include local intelligence to perform a preprocessing of the image. In this case, the scanner will generate bursts of files to the database server. Those files may vary in size from 2 kbyte corresponding to a full ASCII page that has been recognized with an Optical Character Recognition (OCR) mechanism located in the scanner, to a maximum of 40 kbyte, corresponding to a complete page digitized according to the Modified Huffman Code (Group 3-facsimile standard). The compression in this particular case would reach a factor of 10. The servers which are useful to add to the network are: multimedia database (MDB), electronic mail, formatter and printing server. The MDB is a general purpose processor with dedicated hardware to control the access to the disc. The storage medium consists of optical discs (12in disc containing 1 Gbyte and the new 51/4in compact disc with 600 Mbyte capacity), in either single units or grouped in a 'juke-box' of 500, or disc-pack of 50. This new medium will motivate the office system designers to create huge data centres that might even be duplicated (to enhance reliability, or to partition the load among different servers). Workstations will send short queries (10 to 100 byte) and get back small or large files. In some applications, broadcast or multicast can be used to send a long file to a group of users (for example, in education applications, group decision making, etc.). An electronic mail server will store short and long messages. Although present mail systems exchange short and occasionally long textual messages, there are strong reasons to predict that long documents containing Scanner

Electronic

~

Multimedia

Local netwo

Workstations

Figure I.

292

Printing server

A typical office information system

graphics, text, and voice comments will be transferred by mail servers in the future. Thus, the mail servers will act as multimedia database servers with the difference that the messages are stored for a short period of time, e.g. 1 min, 1 day, 1 week. Formatting/printing servers could be single units or be integrated. In some cases, formatting is done by the workstation editor, and pages in their final form are transmitted directly to the printing server which does the rasterization. The assumption made here is that printers in the future will be non-impact and thus will be driven by a graphics controller. The rasterization of one page may take from 2 s if the page content is alphanumeric, to 2 min for a complex page with structured graphics, multifonts text, bitmap or facsimile picture and a form, for a 10 MHz 68OO0-based controller. The page description that has been adopted as the basis of this study is Postscript 6. Printing servers without the spooling function will oblige every workstation to regulate its flow and, hence, create a typical traffic that is analysed in a later section. Printing servers which have the spooling function, and which are equipped with a large capacity disc, generate a different traffic pattern, with long files sent during a short session from the WS to the server. In the case of formatting function located outside the WS, traffic will be exchanged between the two servers (printer and formatter) in a more permanent way than between the WS and the server.

THE LOCAL AREA NETWORKS CONSIDERED Three LANs are compared in this study and all three use baseband transmission, a logical-bus architecture and the Carrier Sense Multiple Access (CSMA) protocol. Ethernet, using CSMA/CD (Collision Detection), is a well known standard I. It transmits at 10Mbit/s using coaxial cable as the physical medium of communication. Starlan 4 is a new CSMA/CD, 1 Mbit/s standard, using a twisted pair of wires. Physically, it is a star that converges to a short bus driven in CSMA/CD mode. It was introduced to create a lower cost than the Ethernet LAN standard, by using a lower speed and more economical cabling system. It uses different CSMA/CD parameters to those used by Ethernet. Appletalk 3,7 is a recent LAN announced by Apple Computer Inc. It uses CSMA/CD (Collision Avoidance) over a twisted pair of wires (300 m maximum length) and its data transfer rate is 230.4 kbit/s. It can interconnect up to 32 stations. Its interfaces are very inexpensive, and the Medium Access Control (MAC) protocol is implemented in software. As it is new, a description of the MAC protocol is in order. The transmitting node uses the physical layer's ability to sense if the channel is idle. If the channel is busy, the node waits until it becomes idle. Upon sensing an idle channel, the transmitter waits for a time equal to the minimum Inter-Dialog Gap (400 ms), plus a randomly generated amount. During this 'wait', the transmitter continues to monitor the channel. If the channel remains idle throughout the wait period, then the node sends a Request-to-Send (RTS) short (9 byte) frame to the intended receiver. The receiver must, within the maximum InterFrame Gap (200 ms) return a Clear-to-Send (CTS) frame to the transmitting node. Upon receiving this frame, the transmitter must, within 200 ms, send out the data frame (600 byte of information at most, plus 9 overhead byte).

computer communications

Since there is no collision detection, as in Ethemet and Starlan, the transmitting node learns of a collision from the RTS/CTS exchange. If no CTS is received, after transmission of RTS, a collision is assumed, and the RTS is retried after adjusting the random time interval. This adjustment follows a linear algorithm that changes the back-off distribution dynamically in response to recent network traffic history.

MODELLING AND SIMULATION There are only a few analytical models of systems of workstations interconnected by LANs. In some cases8 queueing theory and approximations have been used to model a distributed computer system. Recently 9' 10 some new theoretical tools were developed for the exact analysis of LANs with many workstations. In all those cases, however, the queueing models used could not be considered in relation to ISO OSI protocol hierarchy. In order to do so, the only recourse was computer simulation. Several Office Information System (OIS) configurations and protocols using the three LANs mentioned above, were modelled using Queueing NetworkAnalysis Program (QNAP2). QNAP2 (copyrighted by CII Honeywell-Bull and IN RIA) is a package for queueing systems modelling and performance evaluation 11. It incorporates many of the known analytical solutions of queueing networks, and also a discrete-events simulator. It uses an algorithmic language, based on PASCALand SIMULA,and a control language. It is a very powerful and easy to use tool, and has been coded in FORTRAN77for portability. The CSMA/CD and CSMA/CD Medicum Access Control protocols, as with the other higher order protocols modelled, were written in QNAP2 language. The QNAP2 simulation programs that were developed, used the Spectral Method for the 95% confidence intervals estimation and for average performance measures evaluation. The simulations were run on a VAX-11/780 under Berkeley 4.2 Unix. In all cases, we considered 32 stations or servers connected to the I_AN, because of the 32 nodes limit of Appletalk on one hand, and the fact that a majority of departmental or group systems will, on average, have this number of nodes.

OFFICE S Y S T E M CONFIGURATIONS: P E R F O R M A N C E EVALUATION The following OIS applications were considered: • document production: printing text, graphics, images © without formatter printer without spooler printer with spooler © with formatter • document browsing from a multimedia database • electronic mail.

Document production In this application 31 stations transmitting pages of text, graphics, images or combinations of them to a printer server are considered. It is assumed that Postscript coding

vol 10 no 6 december 1987

is used for text and graphics so that up to 4 kbyte/page are transmitted, instead of a bitmap of 1-8 Mbit. In the numerical studies, pages (a combination of text and graphics) are considered to be 3 kbyte long after Postscript coding. Image bitmaps are 96 kbyte page, corresponding to the resolution (1024 x 768) of workstations such as the Buroviseurs. In all cases, it was assumed that the average document generation rate at each workstation was 1 every 7 minutes. This is a realistic figure based on observations of secretaries entering document pages on a word processor.

No formatter case It is assumed that pages are sent to the printer in their final format. No delays for formatting are introduced. Two cases are considered:

a) Printer with no spooling Since in this case the printer server has no spooling function and a hard disc, it is assumed that the server can only receive one page at a time for printing. Thus only single page documents are considered transmitted from the WSs. In all three LAN cases (Appletalk, Starlan, Ethernet), the following higher level protocols were considered and modelled: • Pagesare segmented into frames, transmitted frame by frame, then reassembled at the printer server and printed; • The first RTS (in Appletalk) or data frame (in Starlan and Ethernet), that succeeds in reachingthe printer without a collision reserves the printer for the entire page. The printer server then broadcasts to all stations a control message that enables only one station and disables all others. The successful station, having thus established a session protocol with the server, transfers all frames of a page to the printer. After printing that page the server broadcasts another control message that enbles all stations to compete again for the channel, etc. The waiting time for reenabling is equal to the time needed for finishing printing the page in the server. The WSs generate pages using a Poisson distribution with a mean rate of 1 page every 7 min. No new page can be generated at a WS, while the previous one has not been successfully transmitted over the LAN. Appletalk data frames have an information field of 600 byte, while those of Starlan and Ethemet are 1500 byte long. The frame overheads are 9 and 18 byte respectively. Table 1 gives the percentage network utilization when the 31 WSs transmit pages (a combination of text and graphics) to the printer server (3 kbyte/page). The printer speed is varied from 2-10 s/page. The QNAP2 results correspond to 100 min of actual LAN operation. It can

Table 1.

% LAN utilization (text and graphics)

Printing speed Appletalk (s/page)

Starlan

Ethernet

2 4 8 10

0.1844 0.1984 0.1868 0.1772

0.01654 0.01627 0.01759 0.01773

0.8613 0.8426 0.8830 0.8274

293

Table 2. Printing speed (s/page)

Average delay (s/page); text and graphics Appletalk

Starlan

Ethernet

2

0.3685 + 0.0836

0.2398 + 0.087

0.2133 + 0.08532

4

I .I 01 + 0.3389

1.088 + 0.994

0.8377 + 0.2564

8

9.923 + 3.406

7.308 + 4.060

7.830 + 4.575

13.970 + 4.510

10.300 + 3.709

9.479 + 2.042

10

be observed that in all three cases the LAN utilization is less than 1%. Table 2 shows the network introduced delays (including queueing and transmission time), as a function of printer speeds. The 95% confidence limits are also indicated in each case. It can be observed here that I_AN speed does not play a significant role, as the printer speed controls the page transmission. Table 3 is similar to Table 1, except in the case of all 31 WSs transmitting uncompressed image bitmaps (96 kbyte/ page) to the printer server, at an average rate of I pager per 7 min per WS. It can be observed that Appletalk shows considerable utilization, while Ethemet is still used less than 1%. Table 4 is analogous to Table 2 and gives the average network delay per page, for a network operation of 5 h. 1400 pages were transmitted in each case. It can be noted that, in this case, Appletalk is still not saturated, but

Table 3.

% I.AN utilization (image bitmaps)

Printing speed (s/page)

Appletalk

Starlan

Ethemet

2 4 8 10

26.80 27.18 27.57 25.64

6.170 5.777 5.909 5.649

0.5423 0.5146 0.5188 0.5351

Table 4. Printing speed (s/page)

Average delay (s/page); image bitmaps Appletalk

Starlan

it creates greater delays than Starlan or Ethernet. This should be expected, since its utilization is almost 27%.

b) Printer with spooling The printer server is assumed to have a spooling function and a large disc, hence there is no need for establishing a session protocol between the WS and the server. Stations transmit their frames to the server without any control from the latter. Pages are assembled at the printer server and printed on a FIFO basis. Since the printing delay per page does not affect the I.AN operation, only the case of a printer with speed 10 s/page is considered. Table 5 gives the percentage network utilization, the average network delay, the average printer delay (queueing plus printing), the total delay (network plus printer) and the average disc occupancy for the case of 31 WSs transmitting single page documents of text and graphics (3 kbytelpage). Comparingthese figures with the ones in Tables I and 2 (the case without spooler), it can be observed that: • the network utilization is almost the same: • the network delay is much smaller, but the total delay (including LAN delay and printer delay) is roughly the same in the case of Starlan and Ethernet and only slightly higher in Appletalk; • the average disc occupancy at the printer is small. Table 6 gives similar results in the case of transmitting single pages of uncompressed image bitmaps (96 kbyte/

Table 5. Transmission to printer with spooling (text and graphics) Appletalk

Starlan

Ethemet

% Network utilization

0.8534

0.1832

0.01634

Average LAN delay (s/page)

0.1195

0.0244

0.00240

Average print delay (s/page)

26.890

22.730

22.7300

Total average delay (s/page)

27.010

22.754

22.7324

Average disc occupancy (kbyte)

6

5.2

5.2

Table 6. Transmission to printer with spooling (image bitmaps) Appletalk

Ethemet % Network utilization

25.95

Starlan

Ethernet

5.767

0.5496

2

6.038 +_0.341

1 .I 83 ___0.0593

0.3006 + 0.0322

Average I_AN delay (s/page)

3.774

0.813

0.079

4

8.888 +0.852

2.027 +0.314

0.9125 +0.1701

Average print delay (s/page)

22.870

21.540

29.750

8

70.95 -I- 35.17

9.132 + 1.815

6.2600 + 1.764

Total average delay (s/page)

26.644

22.353

29.829

10

201.60 +51.15

21.38 +10.92

21.490 +7.034

Average disc occupancy (kbyte)

294

163

155

230

computer communications

page) to the printer server at an average rate of 1 page per 7 min per WS. The rasterization time introduced by the printer (possibly having a different resolution than the WS) on the received pages, is taken to be included in the printing delay of 10 s/page. From the point of view of network loading, this case is equivalent to the WS transmitting documents composed of 32 pages of text and graphics (bulk arrivals) at a rate of 1 page per 7 min. Comparing these results with the case without spooling (Tables 3 and 4), it can be observed that:

Table 7. Database browsing (1 page/20 s); pages in text revisable mode

• the LAN utilization is almost the same; • the total delay per page is now much smaller in Appletalk, smaller in Starlan, and about the same in Ethemet; • Ethemet sends pages to the printer too fast and this results in larger queueing delays there; • the average disc occupancy is now fairly high; • only Appletalk shows a major delay improvement, at the expense of the cost of a large disc at the printer with spooling function.

Average LAN 0.11950 delay on _+0.00013 browsed pages (s/page)

Documents sent to a formatter before printing In the case of the WS transmitting pages to a formatter before final printing, something of the order of 1 rain may be added to the delay of each page. This will only affect the LAN performance in the case of a printer server without spooling function. In that case, the earlier results (Tables 1-4) will only be modified in the following way: way: • the LAN utilization will not significantly change; • the average LAN delay per page and the total average delay per page will increase according to the delay introduced by the formatter. If the printing server has a spooling function, the I_AN delay and utilization (Tables 5 and 6) will not be affected, but the total delay per page will increase according to the delay introduced by the formatter.

Sequential browsing from a multimedia database Sending documents to a multimedia database (MDB) with an optical disc introduces a modelling and performance evaluation problem very similar to the one studied above (i.e. document production). As seen in the case of a printer server with spooling, the service time (printing time) did not affect the LAN performance. Thus, if again 31 WSs are considered, sending documents to an MDB server for storage at an average rate of 1 document per 7 min per WS, the LAN utilization and the average I_AN delay per page is given in Tables 5 and 6 (text and graphics or images, respectively). A heavy traffic situation may arise during rapid retrieval of pages from the database. This is the case where one or more WS wish to sequentially browse a file stored in the MDB. If all WS wish to do this, the browsing from the same MDB server takes control of the channel most of the time (one-to-many transmission). The case of one WS browsing, while the others perform document production, is therefore considered. Pages may be stored in the MDB in one of the following modes:

vol 10 no 6 december 1987

Appletalk % Network utilization Average LAN delay on print (s/page)

Starlan

Ethemet

1.035

0.3032

0.03032

0.11960 +0.00013

0.02436 _+0.00013

0.002453 +0.000013

0.02431 0.002448 _+0.000002 _+0.0000013

1) Text revisable form, 2-3 kbyte/page (short time for displaying in the WS); 2) Text revisable form, plus structured graphics (few seconds' delay, for rasterization in the WS) 3) Text revisable form, plus structured graphics, plus image bitmaps at WS resolution (few seconds delay, for rasterization in the WS); 4) Text revisable form, plus structured graphics, plus image bitmaps at higher resolution than in the WS (10-15 s to 1 min rasterization delay in the WS). Because of the large delays in the latter case, browsing is not practical and it is therefore taken that structured graphics and images are stored in the MDB in bitmap mode with the same resolution as the WS. The following OIS scenario is studied: of the 32 nodes connected to the LAN, 29 are WS transmitting pages (text and graphics or image bitmaps) to a printer server with spooling (printing speed 1 page/10 s), at an average rate of 1 pager per 7 min per WS, while a discless WS browses pages (in text revisable mode or bitmap mode) from a MDB. The MDB server is assumed to present 1 page per either 4 or 20 s to the LAN. Table 7 presents the results of the QNAP2 simulation in the case of the 29 WS transmitting pages of text and graphics (3 kbyte/page) to the printer server, and one station browsing pages stored in the MDB in text revisable mode (3 kbyte/page) and emitted at a rate of I page/20 s. The network load presented by the MDB can be observed to be very low. The delays are the same as in Table 5; pages arrive at the browsing station almost synchronously. Table 8 presents similar results for the case of a much

Table 8. Database browsing (1 page/4 s); pages in text revisable mode Appletalk % Network utilization Average LAN delay on print pages (s/page)

3.55 0.1198 +0.00025

Starlan

Ethernet

0.7828

0.07832

0.02440 _+0.00032

0.00245 +_0.000013

Average LAN 0.1195 0.02435 0.00244 delay on +0.000343 _+0.000055 +0.0000002 browsed pages (s/page)

295

Table 9. Database browsing (1 page/4 s); pages in bitmap mode Appletalk % Network utilization

--

Starlan

Ethernet

25.63

2.538

Average LAN delay on print pages (s/page)

0.9669 +0.0373

0.08081 +0.00143

Average LAN delay on browsed pages (s/page)

0.8178 0.07865 +0.008473 +0.00037

faster browsing. It can be observed that all LANs handle this application with very low utilization; pages arrive at the browsing station almost synchronously. Finally, a worst-case traffic scenario is considered, where 29 WSs are transmitting image bitmaps (96 kbyte/ page) to the printer server, while one WS browses pages from the MDB, stored in bitmap mode and at WS resolution (96kbyte/page). The browsing speed is 1 page/4 s. Table 9 displays the results of the QNAP2 simulation in this high load situation. As in Tables 7 and 8, the 95% confidence limits for the average I_AN delays are also given. The following can be observed: • Appletalk saturates and cannot handle this heavy traffic. If the MDB is distributed and the WS has a disc, this difficulty could be overcome by browsing locally. • Starlan has almost 26% utilization and adds almost 1 s delay on the browsed pages. The browsing synchronism is not seriously affected. • Ethemet utilization is still very low, and delays on browsed pages are negligible. This shows again that Ethernet has a very high capacity for non real-time applications in a 'departmental' OIS environment.

Electronic mail In this OIS application, WSs transmit to and, more frequently, receive their mail from, an electronic mail server. A network load peak typically occurs early in the day, during mailbox browsing. From the modelling point of view, this application is no different from the cases modelled and studied above, thus there is no need to model it separately. The traffic here, in general, is not as heavy as in the applications of document production and document browsing, thus the LAN utilization and delays are going to be smaller than those reported above.

CONCLUSIONS The purpose of this study has been to compare three well known LANs (namely Appletalk, Starlan and Ethernet) as to their performance and utilization in carrying typical integrated Office Information Systems applications in a departmental or group environment of around 30 workstations. Because of the necessity to realistically model the Medium Access Control and higher level protocols in

296

the workstations and servers of the integrated OIS, discrete-events simulation was employed as a modelling and performance analysis tool. Both realistic and worstcase traffic scenarios have been considered, in order to obtain upper bounds of the LANs' capacity and performance. The authors have concluded that inexpensive and low data rate LANs, such as the 230.4 kbit/s Appletalk and the faster I Mbit/s Starlan, can easily support typical office applications. Ethernet is not properly utilized in such small group OIS environments, running non real-time applications, and in addition is considerably more costly. In much larger office environments, running real-time (isochronous) applications, such as digital voice and image communications, even Ethernet may become quickly saturated. The new developing LAN standards should then be considered, such as the 43 Mbit/s IEEE802.6 slotted fibre ring12 and the 100 Mbit/s FDDI token-passing fibre ring13. This study has been designed to assist OIS managers in choosing the I_AN most suitable to their needs.

ACKNOWLEDGEMENTS This work was supported by the Bull corporation, the Natural Sciences and Engineering Research Council of Canada, the NATO 'Double Jump' Research Fellowship program and INRIA. REFERENCES 1 Shoch, J F et al. 'Evolution of the Ethernet local computer network' /EEE Computer (August 1982) pp 10-25 2 Shoch, l F and Hupp, J A'Performance of an Ethernet local n e t w o r k - - a preliminary report' Proc. /_AN Syrup. Boston (May 1979) 3 Sidhu, G S, Andrews, R and Harslem, E F 'Applebus: beyond personal computing to personal networking' /EEE 1st Int. Conf. Office Automation New Orleans, USA (December 1984) 4 IEEEDraft Standard P802.3:1 Mbit/sec Starlan. Rev. A, Draft, July 1984 5 Naffah, N Office information systems North Holland, Amsterdam (1981) 6 Postscript language manual Adobe Systems Inc. (August 1984) 7 Inside Appletalk Apple Computer Inc. (1985) 8 Goldberg, A, Popeck, G and Lavenberg, S S 'A validated Distributed System Performance Model' Proc. Performance'83, North Holland, Amsterdam (1983) pp 251-268 9 Balbo, G, Bruell, S C and Ghanta, S 'The solution of homogeneous queueing networks with many job classes' Proc. Int. Workshop Model Perform. Evaluation of Parallel Syst. Grenoble, France (1984) pp 385-417 10 Conway, A E and Georganas, N D 'An efficient algorithm for semi-homogeneous queueing network models' Proc. Performance'86: ACM Sigmetrics '86 Raleigh, NC, USA (May 1986) 11 Potier, D 'New users' introduction to QNAP2', IN RIA Technical Report No 40 (October 1984) 12 Kositpaiboon, R and Georganas, N D 'A performance study of an IEEE 802.6 ring using a simple isochronous management protocol' Proc. 13th Biennia/Syrup. Commun. Kingston, Ontario, Canada (June 1986) 13 Ross, F E 'FDDI-fiber, farther, faster' Proc. Infocom '86 Miami (April 1986) pp 323-330

computer communications