W orkload Generation for ns Sim ulations of Wide Area Net - PDF document

� W orkload Generation for ns Sim ulations of Wide Area Net w orks � and the In ternet y z z y M� Y uksel � B� Sikdar K� S� V astola and B� Szymanski y Departmen t of Computer Science z Departmen t of Electrical Computer and System Engineering Rensselaer P olytec hnic Institute T ro y � NY ����� USA f yuksem�bsikdar�v astola g � net w orks�ecse�rpi�edu� szymansk�cs�rpi�edu Ph� ��������������� � This w ork supp orted b y D ARP A under con tract n um b er F���������C����� � Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y

� Outline of the talk The ns net w ork sim ulator � Sim ulating wide area net w orks� Issues � T ra�c comp osition and proto col di�erences � T ra�c generation � T op ology � Sim ulating wide area net w orks� Solutions � T ra�c comp osition � T ra�c generation � Sim ulation results � Summary and conclusions � Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y

� Sim ulation platform� ns ns� Net w ork sim ulator dev elop ed b y UCB � �LBLN and the VINT pro ject Op en source co de and particularly easy � to mo dify Includes libraries of top ology and tra�c � generators and visualization to ols Wide acceptance in the net w orking re� � searc h comm unit y Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y

� Sim ulating wide area net w orks� Issues T ra�c comp osition and proto col di�er� � ences Sim ulated tra�c should main tain the prop er � comp osition Sources select their destinations randomly � Di�eren t applications generate sessions with � di�eren t distributions Num b er of sessions in a net w ork v aries con tin� � uously is incapable of accoun ting for these e�ects ns � Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y

� Sim ulating wide area net w orks� Issues �Con t�� T ra�c generation � T race based tra�c generators� Unable to ac� � coun t for c hanging net w ork conditions and feed� bac k Source based tra�c generators � do es not use application sp eci�c sto c hastic ns � mo dels to generate tra�c do es not ha v e generators for long�range de� ns � p enden t �Self�similar� net w ork tra�c Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y

� Sim ulating wide area net w orks� Issues �Con t�� T op ology issues � W ANs and the In ternet can b e view ed as a � div erse collecti on of in terconnected domains Eac h domain has its o wn in ternal top ology � Large net w orks also ha v e wide v ariations in � their link bandwidths has libraries for top ology generation ns � Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y

� Sim ulating wide area net w orks� Solutions T ra�c comp osition � W e ha v e implemen ted mec hanisms to con trol � the p ercen tage of TCP and UDP sessions in the tra�c Our implemen tation randomly c ho oses source� � destination pairs from the no des in the net� w ork Capabilit y to use an y suitable distribution to � c haracterize w orkload division b et w een no des and use it to select the source�destination pairs Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y

� Sim ulating wide area net w orks� Solutions �Con t�� Session generation � W e in tro duce application sp eci�c tra�c gener� � ators in for T elnet� WWW� FTP and SMTP ns W e use empirical distributions to c haracter� � ize the in ter�arriv al times� duration and data transferred b y eac h applicati on Sessions generated b y eac h application c harac� � terized b y mean n um b er of sessions �MNS� � mean in ter�arriv al time of sessions �MIATS� � mean duration time of sessions �MDTS� � Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y

� Sim ulating wide area net w orks� Solutions �Con t�� Session generation � Eac h expiring session replaced b y random n um� � b er of sessions Con tin uous v ariation in the n um b er of activ e � sessions giving a more realistic net w ork sce� nario Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y

�� Sim ulating wide area net w orks� Solutions �Con t�� T ra�c generation� Self�similar tra�c gen� � erators Long�range dep endence in aggregated WWW� � Ethernet and W AN tra�c W e implemen ted t w o self�similar tra�c gener� � ators in ns Based on su� Application�Traff ic�Su pFRP� � p erp osition of F ractal renew al pro cesses Based on sup er� Application�Traff ic�SS � � p osition of Mark o v mo dulated P oisson pro� cesses Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y

�� Sim ulation results� Mean duration times of sessions 55 Mean Duration Time (Seconds) 50 45 Telnet WWW 40 FTP SMTP 35 30 25 2 4 6 8 16 10 12 14 Simulated Time (Hours) Mean n um b er of sessions for v arious application � as a function of the sim ulation length The mean duration times of sessions con v erges to � the desired v alue ��� sec� within a short sim ula� tion time Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y

�� Sim ulation results� Mean n um b er of sessions 12 11 Mean Number of Sessions 10 9 Telnet WWW 8 FTP SMTP 7 6 5 2 4 8 6 10 12 14 16 Simulated Time (Hours) Mean n um b er of sessions for v arious application � as a function of the sim ulation length The con tin uous v ariation of the n um b er of ses� � sions results in a closer mo del of real net w orks Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y

�� Sim ulation results� Self�simil ar tra�c generators 1.0 0.8 H = 0.90 H = Covariance (SupFRP) H = 0.75 0.50 0.6 Poisson 0.4 0.2 0.0 -0.2 20 40 60 80 k (lag) V ariance�lag plot for the tra�c generated SupFRP � compared with a P oisson pro cess W e note that the correlations deca y extremely � slo wly and follo w a p o w er la w c haracteristic of self�similar pro cesses Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y

�� Sim ulation results� Self�simil ar tra�c generators 1.0 0.8 H = 0.90 H = H = 0.75 0.50 Covariance (SS) 0.6 Poisson 0.4 0.2 0.0 -0.2 20 40 60 80 k (lag) V ariance�lag plot for the tra�c generated com� SS � pared with a P oisson pro cess W e note that the correlations deca y extremely � slo wly and follo w a p o w er la w c haracteristic of self�similar pro cesses Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y

�� Summary and conclusions W e presen ted a metho dology for gener� � ating realistic w orkloads for W ANs using the net w ork sim ulator ns Our metho d captures the temp oral and � spatial correlation in net w ork tra�c W e created to ols to generate tra�c sp e� � ci�c to applications lik e T elnet� FTP � WWW and SMTP Implemen ted t w o accurate self�similar traf� � �c generators in ns to sim ulate long�range dep enden t aggregate tra�c Departmen t of ECSE Rensselaer P olytec hnic Institute� T ro y