Internet DBMS versus Traditional DBMS Local distributed database - PDF document

Efficient Overload Protection Using SPC Efficient Overload Protection Using SPC Victor Shi Victor Shi Victor_Shi@ndsu.nodak.edu Victor_Shi@ndsu.nodak.edu http://red.atm.cs.ndsu.nodak.edu/vshi/ http://red.atm.cs.ndsu.nodak.edu/vshi/ Internet DBMS versus Traditional DBMS Local distributed database system • Much more users, need high throughput SPC DB node 1 • longer network delay, higher 2 concurrency • Vulnerable to Switch hacker’s attack, and prone to overload. Need overload N protection. The "1999 Computer Crime and Security Survey" found system penetration by outsiders increased for the third year in a row with 30% of respondents reporting intrusions. Those reporting their Internet connection as a frequent point of attack rose for the third straight year, from 37% of respondents in 1996 to 57% in 1999. 1

Solutions to “Denial of Service” “The assaults that battered Yahoo and eBay and a variety of major sites were brutally simple. There are programs that fire off streams of data packets like water from a fire hose”. --Bray: An Assault on The Home Front (2/11/00) Access admission control policies • CS (Complete Sharing) • CP (Complete Partitioning) • TR (Trunk Reservation) • UL (Upper Limit bounds) • GM (Guaranteed Minimum bounds) • UL/GM • FP (full preemption), PP (Partial Preemption) • CnP (Conditional Preemption) More details on the solutions to “denial of service” Access admission control policies • CS (Complete Sharing) • CP (Complete Partitioning) • TR (Trunk Reservation) • UL (Upper Limit bounds) • GM (Guaranteed Minimum bounds) • UL/GM • FP (full preemption), PP (Partial Preemption) • CnP (Conditional Preemption) 2

Access Admission Control Policies: overiew Upper Limit /Guaranteed Complete Minimum Complete Sharing Partition Trunk Conditional Partial Reservation Preemption Preemption Trunk Reservation (TR) 2 units are dedicated to j 1 , 4 TR provides a one- units are shared by j 1 and j 2 . way protection against overloads by rejecting requests of j 2 lower priority when the available resources in system are less than a pre- specified threshold. j 1 3

UL, GM & UL/GM Upper Limit (UL) specifies a maximum 2 units are dedicated to j 1 , amount of resources that each class of 2 units are dedicated to j 2 , requests can use, to limit the overload 2 units shared by j 1 and j 2 within a certain degree. Guaranteed Minimum (GM) reserves a minimum amount of resources for each class of requests to protect it from negative effects caused by overloads of other classes of j 2 requests. UL/GM is a combination of UL and GM , which polices the overload of each class of requests while guaranteeing a minimum performance to it by resource reservation. Thus UL and GM are special cases of the UL/GM j 1 policy. FP, PP and CnP j 2 reclaims its reserved resources CnP allows the reserved (2 units) resources to be shared by requests from all classes of users. When a request j 1 reclaims arrives and cannot find j 2 its reserved sufficient resources, resources preemption is activated (2 units) to revoke its reserved resources used by requests of other classes. j 1 4

Comparison of Control Policies (cont’d) • NEC USA Inc. researchers favor TR S. K. Biswas and B Sengupta,, “Call admissibility for multirate traffic in wireless ATM networks”, Infocom’97, 1997. • AT&T researchers favor UL/GM G. L. Choudhury, K. K. Leung and W. Whitt, “Efficiently providing multiple grades of service with protection against overloads in shared resources”, AT&T Technical Journal, July-August, 1995, pp.50-63. Performance • blocking probability: the probability that a newly arrived service request is rejected for some reason. • preemption ratio: the ratio of the number of requests being preempted in a class to the number of requests being accepted in the same class. • System throughput: the number of requests serviced by the system. 5

Experiments • Symmetric case study In the symmetric case we assume that all classes of users in the system have the same grade of service requirements and workloads. • Asymmetric case study In the asymmetric case requests from different classes have different requirements. System Model We consider a link with two classes of traffic. • A service needs one unit of resource. • Request arrival processes are Poissonian. • The request service times are negative exponential distributed. • The system enforces the access admission control policies by associating appropriate numbers with each class. 6

Symmetric Case Study = = = − = = − 6 × 6 b ( 1 ) b ( 2 ) b 10 , f ( 1 ) f ( 2 ) on the normal workload condition, b ( 2 ) 10 K , = × + − 6 = × f ( 1 ) f ( 2 ) ( 1 x ) on the overload condition 1 and b ( 1 ) 10 K , = × + K denotes the requirement relaxation f ( 2 ) f ( 1 ) ( 1 x ) on the overload condition 2. x denotes the overload percentage. Let K =1, 10, 100 under overload conditions, while and = x 0.5, 1, 1.5, 2, we experiment with 12 scenarios for both UL/GM and CnP policies. = − 6 Table 1: Throughputs of UL/GM policy and CnP policy under b 10 constraint UL/GM Policy K =1 K K =10 =100 Overload ( x ) n 1 , n Throughput n 1 , n Throughput n 1 , n Throughput ( ) ( ) ( ) 2 2 2 0.5 (49,49) 48.867 (45,45) 54.257 (42,42) 57.226 1 (49,49) 48.378 (47,47) 51.523 (45,45) 54.335 1.5 (49,49) 47.773 (48,48) 50.273 (46,46) 53.066 2 (50,50) 47.539 (48,48) 50.273 (46,46) 53.066 = pt = − 7 pt ( 1 ) ( 2 ) 10 CnP policy with preemption constraint requirements Overload( x ) n 1 , n n 1 , n n 1 , n Throughput Throughput Throughput ( 2 ) ( 2 ) ( 2 ) 0.5 (48, 48) 51.894 (48,48) 55.566 (48,48) 59.414 1 (49,49) 48.691 (49,49) 52.822 (49,49) 58.129 1.5 (49,49) 47.441 (49,49) 51.894 (49,49) 57.424 2 (50,49) 47.109 (50,49) 51.601 (50,49) 57.148 = pt = − 6 pt ( 1 ) ( 2 ) 10 CnP policy with preemption constraint requirements x n 1 , n n 1 , n n 1 , n Overload( ) ( 2 ) Throughput ( 2 ) Throughput ( 2 ) Throughput 0.5 (50,50) 53.085 (50,50) 56.992 (50,50) 61.601 1 (50,50) 49.726 (50,50) 54.082 (50,50) 59.619 1.5 (50,50) 48.769 (50,50) 53.369 (50,50) 59.064 2 (50,50) 48.476 (50,50) 53.085 (50,50) 58.789 7

Asymmetric Case Study = − = − = = − 3 × 6 3 b ( 1 ) 10 , b ( 2 ) 10 , f ( 1 ) f ( 2 ) on the normal workload condition, b ( 2 ) 10 K , = × + = − 6 × f ( 1 ) f ( 2 ) ( 1 x ) on the overload condition 1 and b ( 1 ) 10 K , = × + f ( 2 ) f ( 1 ) ( 1 x ) on the overload condition 2. Similarly we experiment with 12 scenarios for the TR, UL/GM and CnP policies (combinations of requirement relaxations K =1, 10, 100 and overloads = x 0.5, 1, 1.5, 2). = − = − 6 3 Table 2: Throughputs of TR, UL/GM and CnP under b ( 1 ) 10 , b ( 2 ) 10 constraints TR Policy K =1 K =10 K =100 ( r ) ( r ) r ) Overload ( x ) Throughput Throughput ( Throughput 0.5 5 55.351 6 60.507 6 68.242 1 5 46.054 5 50.800 5 58.925 1.5 4 39.882 4 43.691 5 50.395 2 4 34.853 4 38.378 4 44.414 UL/GM Policy Overload ( x ) ( n 1 , n ) Throughput ( n 1 , n ) Throughput ( n 1 , n ) Throughput 2 2 2 0.5 (42,55) 56.806 (0,53) 60.117 (0,52) 61.738 1 (43,55) 55.947 (35,53) 58.535 (0,51) 61.972 1.5 (43,55) 55.507 (38,54) 58.476 (25,52) 60.820 2 (44,56) 55.273 (39,54) 58.476 (30,52) 61.738 8

= − = − 7 4 CnP policy with preemption constraint requirements , pt ( 1 ) 10 pt ( 2 ) 10 Overload ( x ) n 1 , n Throughput n 1 , n Throughput n 1 , n Throughput ( ) ( ) ( ) 2 2 2 0.5 (40,56) 58.007 (25,57) 63.945 (20,63) 68.222 1 (42,55) 55.551 (40,56) 61.835 (20,61) 66.799 1.5 (44,55) (41,57) 62.070 (35,62) 67.617 55.244 2 (44,55) (42,56) 61.445 (35,62) 67.617 55.083 = − = − 6 3 ( 1 ) 10 ( 2 ) 10 CnP policy with preemption constraint requirements pt , pt Overload ( x ) n 1 , n n 1 , n n 1 , n ( ) Throughput ( ) Throughput ( ) Throughput 2 2 2 0.5 (41,59) 61.523 (37,63) 67.988 (25,69) 75.239 1 (44,56) 56.914 (42,58) 64.726 (32,68) 74.257 1.5 (44,56) 56.660 (42,58) 64.611 (33,67) 72.382 2 (44,56) 56.523 (42,58) 63.989 (33,67) 70.332 Summary • CnP consistently outperforms TR and UL/GM while the preemption level is acceptably low (Throughput 21% over UL/GM). • We plan to use SPC and PBX card for enforcing CnP policy, so as to efficiently resolve the “denial of service” problem. 9

10