Simulation of a paging computer system*

Transcription

1 Smulaton of a pagng computer system* W. P. Boote.f S. R. Clark.t and T. A. Rourke The operaton of the Atlas computer at Manchester Unversty s smulated usng a next-event type of smulaton model. The agreement between the smulated results and the real system s found to be good. Varous advantages and dsadvantages of ths type of smulaton are dscussed. (Receved May 970) The analyss of the operaton of the Compatble Tme Sharng System (CTSS) by Scherr (967) has been referenced often n the lterature as evdence of the valdty of next-event type smulaton models for the smulaton of computer systems. Many smulatons of computer systems have been carred out but Scherr's s farly unque n that t ncorporates a comparson wth a real system. (Huesmann and Goldberg, 967; Estrn and Klenrock, 967.) Events are created by jobs fnshng, by tme slces beng fnshed, and by the start and fnsh of transfers to the consoles or the drum. Events whch occur on a much shorter tme scale, for example addtons, multplcatons, memory accesses, etc., are gnored. It s the success of the smulaton n the lght of the latter assumpton whch makes the work so nterestng. The CTSS system, however, dd not overlap the operaton of the drum/dsk and the processor, and there was no need to represent the sze of the core memory n the smulaton (Scherr, 967). It s nterestng to consder what would happen f Scherr's type of comparson was carred one stage further to nclude a fnte core, and the overlap of processng wth tape and drum transfers. In ths stuaton, gnorng the effects of memory accesses s less lkely to be successful. Such a comparson s the purpose of ths artcle. Systems whch ncorporate pagng are troublesome from a smulaton pont of vew snce the page-turnng events take place much more frequently than progam-swappng events on a non-paged computer. The real tme necessary to complete a smulaton s therefore longer. Scherr's smulaton had a smulated tme to real tme rato of approxmately usng an IBM 709 (Scherr, 967), whereas Nelsen's smulaton of an IBM 60/67 on an IBM 60/50 had a rato close to (Nelsen, 967). Snce pagng systems are more complex than non-pagng systems and troublesome from a smulaton pont of vew, we feel that a comparson between smulaton and practce for a pagng system would be of wder nterest than for a non-pagng system. In addton, pagng has become an accepted manner of accessng nformaton. Statstcs concernng the operaton of the Atlas computer at Manchester Unversty are avalable n the lterature (Morrs, Sumner, and Wyld, 967), and ths system seems a very reasonable choce. We were able to obtan addtonal statstcs from the Atlas loggng tapes whch were made avalable. A further advantage n the choce of the Atlas system was apparent when t was found to be possble to solate the nner schedulng for smulaton, and gnore card readers, paper tape readers, prnters and punches. Ths s possble because jobs are tmed by the Atlas system between readng n and prntng out. The Atlas system The purpose of ths secton s not to descrbe fully the operaton and schedulng of Atlas but rather to brng out the salent features necessary for an understandng of the smulaton model. Full detals of the Atlas system are already present n the lterature (Morrs et al., 967; Howarth, Jones, and Wyld, 96; Klburn et al., 96.) The basc queueng whch takes place n the Manchester Unversty Atlas s shown n Fg.. Jobs queue for the executon phase after beng read n (nput well). The number of jobs smultaneously n the executon phase s lmted to two by the supervsor and t may be further lmted by the sze of the core and the drum. The core holds pages of nformaton and the drum holds a further 9 pages, of whch are avalable to a job. Between the stages of queueng for the executon phase and beng accepted for executon there s an addtonal stage of assembly. Assembly nvolves locatng the pages of a job (wth, any necessary complers) on the drum, transferrng the frst page to core and mountng ts user tapes (Atlas has eght magnetc tape decks whch can be assgned to jobs). An assembled job s therefore ready to begn executon wthout further delays. The supervsor assembles jobs accordng to a prorty scheme whch tres to mantan a tape job and a non-tape job n the executon phase at the same tme (Internatonal Computers Lmted, 96). In the executon phase pages are transferred to or from drum, tape transfers may be made to one or more of the magnetc tapes assgned to a job, and output may be created on the output well located on dsk. No queueng takes place for transfers to or from user tapes because each tape s connected to core va a separate channel. The queue dscplne for drum transfers and for the use of the output well s frst-come-frst-served. In the queue for CPU attenton, however, tape jobs are gven prorty over non-tape jobs. A subdvson of the category of non-tape jobs s that of short jobs and long jobs. The dstncton between these two s that long jobs request n excess of 0 seconds of CPU tme. Tape transfers are freely overlapped wth drum transfers and CPU executon. However, CPU operaton s not overlapped wth drum transfers nto core. The transfer tmes and overheads for the system (Morrs et al., 967) are: mean tape transfer tme = 006 seconds/page mean drum transfer tme = 00 seconds/page (swappng a page) supervsor or overhead to transfer control to a job = 000 seconds, supervsor overhead to locate a page on drum = seconds. Downloaded from by guest on 0 December 08 Ths work was carred out at the Unversty of Mantoba, Wnnpeg, Canada, wth support from the Natonal Research Councl of Canada jpresent address: Internatonal Ar Traffc Authorty, Geneva, Swtzerland ^Present address: Department of Mathematcs, Unversty of Vctora, Vctora, Canada Present address: Techncal Support Unt, Department of Trade and Industry, London Volume 5 Number 5

2 ST/ RT OF SYSTEA«\ TIMING END OF SYSTEM TIMING ARRIVALS JOBS WAITING IN THE INPUT WELL ASSEMBLED JOBS OUTPUT WELL Fg.. The basc queueng n the Manchester Unversty Atlas Atlas statstcs The loggng nformaton from sx weeks of operaton of Atlas was processed n order to obtan nformaton charactersng the Atlas job stream (to be nput to the smulaton model) and Table Atlas nput statstcs for the sx-week perod commencng 5 September 968 (a) (b) (c) * (d) CLASS NO. j COMPUTE TIME RANGE (sees.) < < < < ; (sees.) "j INPUT STATISTICS REQUESTED (pages) COMPILE nformaton charactersng the treatment gven to ths job stream by the Atlas schedulng (to be compared wth the output of the smulaton model). These are referred to as nput and output statstcs. STORE EXEC MEAN NO. OF TAPES MEAN NO. OF PAGES TRANS.* PRINTER OUTPUT (lnes) (sees.) Downloaded from by guest on 0 December 08 (<0 < f) < Rewnd tme converted to an equvalent number of pages transferred 5 The Computer Journal

3 Table Atlas output statstcs for the sx-week perod commencng 5 September 968 (a) (*) (c) (d) (e) CO CLASS ATj Wj NO. (sees.) (sees.) j Tables and show the nput and output statstcs for the sx-week perod 5 September to 6 October 968. Jobs are separated nto the same compute tme ranges that have been consdered prevously, Morrs et al. (967), and,-, n t and dj are the mean compute tme, the mean number of jobs n one week of Manchester Unversty operaton and the mean drum tme (for pagng), for each compute range j. After the complaton phase s complete the supervsor reserves the amount of store requested for the executon phase for the remander of tme. The rather large value of the observed number of blocks transferred to or from user magnetc tapes s a result of convertng the tape rewnd tmes attrbuted to the jobs to an equvalent number of tape transfers. The output statstcs consst of two performance characterstcs. The quantty ATj s smply the mean elapsed tme between the start and fnsh of the system tmng of a job (see Fg. ) calculated over each compute range / Each job s tmed between the events of the end of assembly and the end of the executon phase (ncludng transfer of nformaton to the output well). The quantty U5 } s the elapsed tme multplcaton factor (ETMF) calculated for compute range j (Stmler, 969). The ETMF of job of compute range j s gven by The smulaton model The general flow-dagram of the smulaton model s shown n Fg.. The event-control routne determnes when jobs arrve, when tape transfers are requested, when page transfers are requested, etc., usng a random number generator. Each newly generated event (ntal event) s recorded n a table along wth the termnal events such as the completon of a page transfer, completon of a tape transfer, completon of assembly, etc. Ths table s scanned on each pass through the routne n order to determne the next event. (Benefts may arse from havng ths lst arranged n tme sequence (Ulrch, 968).) The event-control routne also contans the Atlas logc for the selecton of jobs for the assembly phase. The search routne searches the queue of jobs awatng CPU attenton n order to fnd a job for executon. Ths search s also carred out usng the Atlas logc to determne whch jobhas prorty and when the CPU executon may overlap current transfers. If a job s found for executon the servce routne s entered to engage the CPU and to record the tme when the job would fnsh f gven unnterrupted CPU attenton. Ths tme s posted n the next-event table for consderaton by the event-control routne as the next event. Whenever the processng of a job s nterrupted by the occurrence of an event, the update routne determnes the amount of CPU tme remanng before the job can fnsh. It also records the amount of tape tme and drum tme used. Whenever a job has fnshed executng and ts output has been transferred to the output well, the removal routne s entered to record the features of the job (CPU tme requested, number of tapes, etc.) and the treatment receved by the job (response tme, ETMF and the queue tme). The storage space and user magnetc tapes assgned to the job are then released. The nput statstcs from Atlas whch were nput to the model are used by the event-control routne to determne the detals of the job stream. The actual statstcs used are the nput statstcs shown n Table \{d). In order not to overcomplcate the smulaton model varous approxmatons were made concernng the dstrbutons of the varates appearng n the nput statstcs. These are dscussed n the Appendx under headngs whch correspond to headngs n Table. Only demanded pages are consdered n the smulaton, that s requests for pages not avalable n core, and the detaled pattern of address (page) references whch occur at the nstructon level are gnored. The nterarrval tme of page demands s assumed to be exponentally dstrbuted wth a mean calculated from the mean drum tme of each compute range, d }. Two approxmatons were employed n the calculaton of the nterarrval tmes of page demands The smpler of these approxmatons gnores the number of pages of a job currently occupy- Downloaded from by guest on 0 December 08 where AT } s the elapsed tme of job /, d t and t, are the compute tme, drum tme and tape tme of job. Clearly ETMF values are normally greater than unty as a result of contenton between jobs wthn the system. Fg.. A generalflow-dagramof the smulaton model Volume 5 Number 5

4 ng the core and assumes that the mean nterval of executon tme between page demands, I s, s constant and gven by Xj = 0-0 ljd } () where the constant 00 seconds s the page swap tme. The Manchester Unversty Atlas has a rather small core and t s nearly always necessary to transfer a page out after a demanded page has been transferred n (Morrs et al., 967). The second of these approxmatons attempts to take nto account the varaton of the number of pages of a job n core. Belady and Kuehner (969) have gven the relatonshp between Xj and the proporton of the pages of a job n core, P o, for completely random page references as -P o ) () and the value of kj s selected here such that the average executon nterval corresponds to the Atlas statstcs assumng that on the average each job occupes half the core. In addton to usng ths equaton for the determnaton of the mean value, lj, the second pagng approxmaton takes nto account to some extent the effect of the varaton of P o between page demands. The event control routne generates the tme of the next page demand for a job as soon as the smulated clock passes an exstng next page demand. Ths generaton s based upon the evaluaton of equaton () and selecton of a value from an exponental dstrbuton wth approprate mean. Between the tme of generaton and the tme of selecton of the page demand the value of P o may change due to pages of a job beng transferred out when the space s requred n core. Ths gves rse to a seres of proportons P o, P t, P,... whch arses at tmes t 0, t u t,.., t 0 beng the tme of generaton of an executon nterval X*. A recursve formula was used to modfy X* to a value approprate to a new proporton > t at tme t u gvng X*. The formula used s ' = wth k =,,... and < () In order to mplement ths scheme n the smulaton model t was necessary to keep track of whch jobs the pages n core belonged to, and to also record all the dynamc changes of state. For the smpler approxmaton (equaton ()) ths was not necessary. Whenever the core becomes full on Atlas as a result of transferrng a demanded page, the drum-learnng program s used to select a page for removal from core n order to mantan a vacant page. Usng nformaton recoded about the actvty of the pages n core, the drum-learnng program uses an algorthm whch seeks to remove the page least lkely to be referenced agan. Ths algorthm s desgned to detect loops. Ths logc s smulated n the model n an approxmate way, and wthout the vast amount of detal necessary to do a complete smulaton of the pagng behavour. Page numbers are generated by roundng values from a truncated normal dstrbuton wth a mean of RJ, a standard devaton of /? y / and a range of 005 to R^ + 0-9, where R^ s the sze of job from compute range j. The page correspondng to the page number generated for a job s brought n for the job when the event s selected by the event-control routne. Snce each page reference s not generated, however, t s only possble to determne whch page s removed n an emprcal way. When space n core s requred the page transferred back to drum s the page wth a page number lyng n the most mprobable regon of any of the dstrbutons of page references for each job on the executon lst. In ths manner a pattern of page usage s establshed clustered about an arbtrary mean of a truncated normal dstrbuton. Results The object of the runs of the smulaton model s to see how closely the output statstcs of the smulaton model match the output statstcs shown n Table. Smulatons based upon the generaton of random events produce results whch have doubtful reproducblty (Abate, Dubner, and Wenberg, 968; Fshman, 967). It s necessary to treat or avod ths problem adequately before the relablty of the results can be determned. We have not determned the reproducblty of the smulaton results obtaned here but nstead we have attempted to show that the results are farly reproducble when compared wth the range of fluctuaton of the output statstcs from the Atlas loggng nformaton. One run of the smulaton model was carred out usng the pagng approxmaton expressed by equatons () and (). The Table Results of the smulatons CLASS NO. j C 00 )) 55 J) 0 )) )* J) 0- S) 0- }) j j» 0-8 a 0 a )) S) -0 JJ ATj (sees.) (Oj (sec; COMMENTS Smpler pagng approxmaton (usng equaton () for the pagng approxmaton) Other pagng approxmaton (usng equatons () and () for the pagng approxmaton). Downloaded from by guest on 0 December 08 5 The Computer Journal

5 C = 0.6 AT; 0 -. A C= n Fg.. Varaton of the ETMF wth the number of jobs for the worst and best cases output statstcs produced can be compared drectly wth Table. The smulaton results are shown n Table. A seres of runs usng the smpler pagng approxmaton (equaton ()) were carred out varyng the pagng rate from zero up to the Atlas fgures. The drum tme was taken as Cdj for each compute range j, where C s a varable factor and dj (j =,,, ) s the drum tme recorded n Table l(a). Values of 00, 0, 0-, 0-,, 0-8 and 0 were adopted for C. Dfferent seres of random numbers were used n each case. A varatonal scheme of ths nature gves a very good dea of the reproducblty of the results. However, ntally the decson was made n the belef that the real tme necessary to run the smulator ncreased wth C n a geometrc manner. (On the bass of all the runs made t was later determned that the ncrease was n fact very close to lnear.) By observng the varaton of the output statstcs ATj and Wj wth n (the number of jobs processed) t was observed that the values always became statonary n the regon n = 00 to n = 500. Fg. shows the varaton of a^ and G5 wth n for C =. These are examples of the worst and the best cases. The results of the smulaton runs whch are shown n Table each nvolve the smulated processng of between 600 and,000 jobs. Fgs. and 5 show how the results for ATj and w } vary wth C. Fgs. and 5 can be used to gan an dea of the reproducblty of ATj and aj values. The apparent ranges compare very favourably wth the values obtaned n the drect calculaton usng the more complex pagng approxmaton. Ths suggests that the smpler pagng approxmaton s as effectve as the more complex approxmaton, and gves further confdence n the reproducblty of the results. Fgs. 6 and 7 show how the calculatons usng the two pagng approxmatons compare wth the Atlas output statstcs recorded n Table. Fg. 6 shows the varaton of ATj wth y (the mean CPU tme request of compute range j) and Fg. 7 shows the varaton of ajj wth,-. It can be seen that n both cases the agreement s good for %j values greater than second. For,- values less than second the scatter of the observed (Atlas) values of AT } s very great and the smulated values le wthn ths range. For Zj values less than second agan the scatter of the Fg.. Varaton of the mean elapsed tme wth the pagng rate factor observed (Oj values s also very great (60 to 60) but ths tme the smulated values are low and do not le wthn the range. Snce the smulated J7\ values le wthn the range of the observed AT values whereas the co, values do not, t seems that t s the components of the denomnator of a> (equaton (I)) whch are at fault. The compute tme, tape tme and drum tmeare added together to form the denomnator. The smulated tape transfers and drum transfers have exponentally dstrbuted nterarrval tmes, and the compute tmes come from a- truncated exponental dstrbuton (see Appendx). If the occurrence of very small denomnators n the Atlas statstcs were more frequent than predcted by the model, t seems qute lkely that rather hgher w, values than the ones predcted would result n practce. Ths s farly reasonable snce t s hghly probable that compute tmes, tape tmes, and drum tmes would be postvely correlated. (No specal correlaton was assumed n the smulaton.) The agreement between smulaton and the actual results observed n Fgs. 6 and 7 s really rather surprsng consderng the number of smplfyng assumptons employed n the smulaton. It s further strong evdence that smulatons of computer systems carred out at ths level of detal can produce relable results. In partcular t appears that memory cycles can be gnored n a machne wth overlapped nput-output and executon. Rather severe smplfyng assumptons concernng the dstrbutons assocated wth the nput means appear to be qute acceptable, n general, but the assumptons for compute range jobs appear to have gven rse to some dscrepancy n S but not n AT. Value of smulaton The value of smulaton appled to computer systems les n ts ablty to provde nformaton about rather complcated systems whch cannot be analysed by queueng theory or Markov technques. A queueng arrangement whch cannot be treated by ether method s present n the Atlas system. Jobs awat executon n the assembly queue and n the prevous watng queue (nput well). The mean tme whch jobs spend n these queues s dependent upon the mean tme taken to process jobs n the executon stage where CPU attenton and nput-output transfers are overlapped n a way whch s characterstc of Atlas. Downloaded from by guest on 0 December 08 Volume 5 Number 55

6 \ C Fg. 5. Varaton of the ETMF wth the pagng rate factor If the number of jobs whch could execute smultaneously were unlmted or lmted to, Markov technques or queueng theory would be applcable. la stuatons where analytcal methods are approprate these methods can be employed more satsfactorly than smulaton methods (Abate et al., 968). However, where analytcal methods are napplcable to part of the queueng structure of a system, smulaton methods should be used for at least that part, replacng other smpler sectons by analytcal methods possbly. Smulaton of a queueng system such as the one n Atlas can be carred out farly rapdly provdng the pagng actvty s not too hgh (n ths context the Atlas pagng actvty s consdered hgh snce no hgher rate was treated). The average tme taken by the Atlas smulaton model relatve to the correspondng tme for a system wth no pagng s shown n Fg. 8, as a functon of C. For the complete Atlas smulaton (C = 0) the observed rato of tmes s approxmately 0. The tme taken to smulate the processng of,000 jobs n a system wth no pagng s about mnutes of IBM 60/65 tme. Ths tme may be reduced to some extent by orderng the nextevent queue n tme sequence. As computer systems become more and more complcated, smulaton models of a complete system become ncreasngly mpractcal. Subdvson of the total system nto sectons for smulaton appears to be the only vable soluton. The smulaton presented here s an example of a central secton. The smulaton does not nclude, for example, card readers and prnters (the outer secton) but n many cases these may be studed separately by whatever means s most convenent. Also the model does not nclude the study of pagng n detal. Approxmately % of all page requests on Atlas are for pages n core (Morrs et al., 967) and n order to study the effcency of pagng algorthms t would be necessary to consder all page requests whether for pages whch are n or out of core, and consderably more detaled nformaton about each job would be necessary to generate these requests. Clearly ths level of smulaton could not be ncluded wthn the present model wthout ncurrng run tmes ncreased by a very large factor. A separate study of pagng at ths level (an nner secton) would obvously be one advantage, and such studes have been carred out n practce (Coffman and Varan, 968). In ths way output from one study could form nput to another. Acknowledgements We wsh to thank the Natonal Research Councl of Canada for grants n support of ths work, the Unversty of Mantoba for provdng tme on ther IBM 60/65, and Mr. G. Rdng for producng the Atlas statstcs from logged nformaton. We are ndebted to the referee for helpful comments and crtcsms of ths work. Appendx A number of smplfyng approxmatons or assumptons were made concernng the dstrbutons assocated wth the nput means whch are used n the model. These nput means appear n Table \{a) and the assumptons are dscussed n the same order. Compute tmes (^y) The compute tmes for each compute range j are selected from a truncated exponental dstrbuton wth a mean of,- (gven n Table \{a)). The exponental dstrbuton from whch the (Upvalues are generated s bounded by the compute range formng a lower and upper lmt to y. Interarrval tmes of jobs The dstrbuton of nterarrval tmes wthn each compute range s assumed to be exponental. The means of the dstrbutons are calculated from the approprate values of rj, the number of jobs processed n compute range j over a perod of one week (Manchester Unversty usage). Requested store {R,j) No nformaton was gathered about the relatve proportons of comple and execute tme and these were not consdered as separate stages n the model. The approxmaton was made that the mean store request was the average of the comple and execute values for each compute range, Rj. Storage requests, R t j, were generated by roundng values from a normal dstr- 000 AT: 00 KEY O SIMPLER PAGING APPROX. a OTHER PAGING APPROX. ATLAS RESULTS 0. 0 _ Fg. 6. Varaton of the mean elapsed tme wth the mean CPU request tme f Downloaded from by guest on 0 December The Computer Journal

7 IOUO KEY 00 O SIMPLER PAGING APPROX. A OTHER PAGING APPROX. ATLAS RESULTS Fg. 7. Varaton of the ETMF wth the mean CPU request tme buton wth a mean of Rj and a standard devaton of Rj/. Only those values of ^satsfyng the condton < R tj < were accepted. Requested magnetc tapes (U tj ) The number of user magnetc tapes requested s taken to be geometrcally dstrbuted about the mean for each compute range, Vy. u =k-\)={\-a)o k ~ l and a = e- /u j neglectng those values of U tj greater than 8. The nterarrval tme of page transfer requests to or from magnetc tapes The tape rewnd or search tme on Atlas s converted to an equvalent number of tape transfers for the statstcs of Table. The average tme taken by a tape for the completon of a sngle request s n the range of three to seven page transfers (Morrs et al., 967). In the model, pages were assumed to be transferred to or from magnetc tape n blocks of eght. The executon nterval of transfers to or from magnetc tape are taken to be exponentally dstrbuted, and the mean rate of occurrence of requests s lnearly related to the number of tapes C Fg. 8. Varaton of the smulaton tme rato wth the pagng rate factor assgned to a job. The mean executon nterval of tape transfer requests, M Jt s. gven by 8, E (k-l)pr(u J = k-l) Mj{m) = -JZ m h where m s the number of user magnetc tapes assgned to a job, and Pj s the mean number of page transfers to or from magnetc tape requested by jobs n compute range/. Number of output pages requested (0 y ) The number of output pages requested s calculated by roundng a varate from a normal dstrbuton wth a mean of Bj and a standard devaton of dj/. d } s calculated by takng the mean number of output lnes per page to be 6 (a fgure used at Manchester). Any value less than was converted to. The output of nformaton to the output well was smulated as a sngle transfer requested at the tme of completon of the processng of a job. References ABATE, J., DUBNER, H., and WEINBERG, S. B. (968). Queung analyss of the IBM dsk storage faclty, JACM, Vol. 5, pp BELADY, L. A., and KUEHNER, C. J. (969). Dynamc space-sharng n computer systems, CACM, Vol., pp COFFMAN, E. G., and VARIAN, L. C. (968). Further expermental data on the behavour of programs n a pagng envronment, CACM, Vol., pp ESTRIN, G., and KLEINROCK, L. (967). Measures, models and measurements for tme-shared computer utltes, Proc. nd ACM Natl. Conf., pp FISHMAN, G. S. (967). Problems n the statstcal analyss of smulaton experments: the comparson of means and the length of sample records, CACM, Vol. 0, pp HOWARTH, D. J., JONES, P. D., and WYLD, M. T. (96). The Atlas schedulng system, The Computer Journal, Vol. 5, pp. 8-. HUESMAN, L. R., and GOLDBERG, R. P. (967). Evaluatng computer systems through smulaton, raecom/wter/ohrwa/, Vol. 0, pp. 56. Internatonal Computers Lmted. (96). The Atlas I computer system operators manual part one, central machne and supervsor, CS. KILBURN, T., HOWARTH, D. J., PAYNE, R. B., and SUMNER, F. H. (96). The Manchester Unversty Atlas operatng system Part I: Internal organzaton, The Computer Journal, Vol., pp. -5. KILBURN, T., PAYNE, R. B., and HOWARTH, D. J. (96). The Atlas supervsor. Proc. AFIPS, Vol. 0, pp MORRIS, D., SUMNER, F. H., and WYLD, M. T. (967). An apprasal of the Atlas supervsor, Proc. nd ACM Natl. Conf., pp NIELSEN, N. R. (967). The smulaton of tme-sharng systems, CACM, Vol. 0, pp SCHERR, A. L. (967). An analyss of tme-shared computer systems, Research Monograph No. 6, MIT Press, Cambrdge, Mass. STIMLER, S. (969). Some crtera for tme-sharng system performance, CACM, Vol., pp ULRICH, E. G. (968). Seral/Parallel event schedulng for the smulaton of large systems, Proc. rd ACM Natl. Conf., pp Downloaded from by guest on 0 December 08 Volume 5 Number 57