U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science Resource containers: A new facility for resource management in server systems G. Banga, P. Druschel and J. Mogul Rice University Presented by: Bhuvan Urgaonkar

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 2 Motivation Web servers should be able to Provide resource guarantees Counter DoS attacks Do accurate billing Shortcomings in OS abstractions Process is unit of resource management Apps have no control over resources kernel consumes for them => achieving above goals difficult

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 3 Outline Motivation Shortcomings of OS Abstractions Resource Containers Performance Evaluation Conclusions/Comments

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 4 Resource Principals and Protection Domains Resource Principals Entities for which separate resource allocation and accounting done Protection Domain Entities that need to be isolated from each other Popular Operating Systems Process = resource principal Process = protection domain

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 5 A classical application Process appropriate resource principal

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 6 A network-intensive application Resources consumed by kernel unaccounted => process inappropriate as resource principal

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 7 A multi-process application The resource principal should be the set of all processes

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 8 Single-process MT application Correct resource principal smaller than a process

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 9 OS shortcomings: summary OS unit of resource mgt. (process) often different from desired unit Apps cannot control resource allocation to their activities Inaccurate accounting and charging of resources Incorrect scheduling decisions

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 10 Resource container Contains all system resources used for an independent activity E.g. For a HTTP connection served by a web server CPU time spent on the connection sockets, PCBs, network buffers etc Mechanisms for accurate accounting Lazy Receiver Processing (LRP)

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 11 Containers and CPU scheduling Applications associate containers with independent activities Dynamic binding between threads and containers Thread’s consumption charged to right container

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 12 Containers and CPU scheduling Threads scheduled based on combined allocation and usage of all associated containers Apps can associate scheduling info with activities Threads within a container may be scheduled as the app pleases Resource container hierarchy

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 13 Other resources Similar accounting mechanism possible for other resources Support required to account for resources consumed by the kernel on behalf of applications

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 14 Containers in a MT server New container created for new connection Serving thread bound to this container

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 15 Containers in event-driven server New container created for new connection Thread’s binding changed as it serves different connections

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 16 More examples Different priorities to requests from different sources Restrict resource consumption of certain requests Defend against DoS attacks Generating accurate bills

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 17 Prototype implementation Modifications to Digital UNIX 4.0D CPU scheduler TCP/IP subsystem: LRP Server software: single-process, event-driven Clients used the S-Client software

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 18 Costs of new primitives Throughput of server unchanged on modified kernel

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 19 Prioritized handling of clients Number of concurrent low-priority clients

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 20 Controlling resource usage of CGI processing Number of concurrent CGI requests

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 21 Immunity against SYN-flooding SYN-Flood Rate (1000s of SYNs/sec)

U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science 22 Conclusions Resource container, an OS abstraction to explicitly identify a resource principal Prototype implementation in Digital UNIX Combined with accurate resource accounting (e.g., LRP) can help web servers provide differentiated QoS