A Workload-oriented Programming Model for Temporal Isolation with VBS Ana Sokolova Department of Computer Sciences University of Salzburg joint work with Silviu Craciunas and Christoph Kirsch

It is about ● Scheduling processes in temporal isolation ● Using variable bandwidth servers for predictability ● Server design for performance the time it takes to execute a piece of code is bounded, independently of concurrently running code

Process model process time action α 1 response time load arrivaltermination action α 2 ● action is a piece of code ● process is a sequence of actions speed

Problem process 1 process 2 process n schedule the processes so that each of their actions maintains its response time uniprocessor Solvable with variable bandwidth servers (VBS) Results: ● a constant-time scheduling algorithm ● a constant-time admission test

Resources and VBS π1π1π1π1 λ1λ1λ1λ1 time π2π2π2π2 λ2λ2λ2λ2 ● VBS is determined by a bandwidth cap (u) ● VBS processes dynamically adjust speed (resource) λ 1 /π 1 ≤ u and λ 2 /π 2 ≤ u ● generalization of constant bandwidth servers (CBS) [Abeni and Buttazzo 2004] virtual periodic resources virtual periodic resources period π limit λ utilization λ/π

One process on a VBS response time process time process running on a VBS response time under VBS π1π1π1π1 time load action α 1 response time action α 2 load λ1λ1λ1λ1 λ2λ2λ2λ2 π2π2π2π2 response time under VBS

VBSprocess running on a VBS time blocked ready runnin g new action preemption due to limit λ preemption due to release dispatch release next action completion response time under VBS arrival termination

VBS process 1 (2,4) process 2 (2,8) process 3 (1,6) multiple processes are EDF-scheduled blocked readyrunning

Scheduling result and bounds For any action α on a resource (λ,π) we have upper response time bound load / λπ + π - 1 lower response time bound load / λπ jitter jitter π - 1 Processes P 1,P 2, …,P n on VBSs u 1,u 2, …,u n, are schedulable if ∑u i ≤ 1

Implementation ● constant-time scheduling algorithm ● different queue management plugins n – number of processes t – number of time instants trade-off time and space

loop { int number_of_frames=determine_rate(); allocate_memory(number_of_frames); read_from_network(number_of_frames); compress_data(number_of_frames); write_to_disk(number_of_frames); deallocate_memory(number_of_frames); } until (done); Programming model loop period action 1 different throughput and latency requirements for different portions of code action 2

How do we get VBS parameters for an action? “server design problem”

Response-time function

f R : R D → Q +

Execution-time function f E : E D → Q + load

Utilization f U (w) = (f E (w) – d E ) / (f R (w) – d R ) f U (w) < c U

Timing of the allocate_memory action

Response-time sampling Scheduled response time (under VBS)

Server Design Problem Finding the right λ,π is difficult. For f S (w) ≤ f R (w) one can choose π as follows: ● 0 < π < d R – d E / c U ● π divides d R evenly ● π divides f R (w) – d R evenly or ● λ divides f E (w) – d E evenly λ = π * c U

Server Design Problem Higher-level scheduler: ● small period for the first part of an action ● large period for the remaining part Smallest π possible: ● f S approximates f R best ● less response-time jitter ● increased scheduler overhead Scheduler overhead accounting: ● utilization accounting ● response-time accounting ● combined accounting

Conclusion For scheduling processes in temporal isolation: ● Programming model as a link to VBS ● VBS provide predictability ● Server design for better performance