Applicability of Reliable Server Pooling for Real-Time Distributed Computing

Reliable Server Pooling defines protocols for providing highly available services. The services are located in a pool of redundant servers and if a server fails, another server will take over. The only requirement put on these servers belonging to the pool is that if state is maintained by the server, this state must be transferred to the other server taking over. The goal is to provide server-based redundancy. Transport and network level redundancy are handled by the transport and network layer protocols. The application may choose to distribute its traffic over the servers of the pool conforming to a certain policy.

The scope of this document is to explain the way of using Reliable Server Pooling mechanisms to manage and access pools of Distributed Computing resources.

The terms are commonly identified in related work and can be found in the Aggregate Server Access Protocol and Endpoint Handlespace Redundancy Protocol Common Parameters document ietf-rserpool-common-param

The application scenario for Distributed Computing is defined as follows: Clients generate large computation jobs. Jobs have to be processed by servers as soon as possible (real-time), i.e. unlike concepts like SETI@home , it is not possible to let clients fetch a job, process it later and may be some day upload the result. Jobs may be partitionable, i.e. they can be split up to smaller pieces which can be processed independently and the processing results can be concatenated to the processing result of the complete job. Jobs have to be processed by servers. Servers may be unreliable; i.e. user computers may be temporarily added to the pool of computing resources and may be revoked when they are used again by their owners. Furthermore, they may simply disappear because of broken network connections (modems, etc.) or power turned off. The processing power of servers in a pool of computing resources may be very heterogeneous, i.e. a few supercomputers and many low-end user PCs. Maintaining a Distributed Computing pool for the scenario described above arises the following requirements to the pool management: It must be possible to manage large server pools, e.g. up to some hundreds or even thousands of servers. Due to heterogeneous processing resources within a pool, it must be possible to use appropriate server selection procedures to meaningfully utilize the available resources. It must be possible to dynamically add and remove servers. Servers may be unreliable, especially when the servers are represented by user PCs. Failover mechanisms are required to continue an interrupted computation session.

All requirements for pool and session management of the Distributed Computing scenario defined in the previous section can be fulfilled by the Reliable Server Pooling architecture: An efficient implementation of the handlespace management structures allows pools to contain thousands of elements. Handlespace management structures have been proposed, implemented and analyzed in , . RSerPool allows to specify server selection rules by pool member selection policies . A set of adaptive and non-adaptive policies is already defined. To fulfill the requirements of new applications, it is also possible to define new policies. Research has already been made on the subject of load distribution efficiency of pool policies in Distributed Computing scenarios: see , , , , for details. Dynamic addition and removal of PEs is a feature of RSerPool . The control/data channel concept of RSerPool realizes a session layer. That is, RSerPool already handles the main task of maintaining and monitoring connections between PUs and PEs; the only task of the application layer to provide full failover functionality is to realize an application-dependent failover procedure. By the usage of client-based state synchronization , in the form of ASAP Cookies, a failover may be fully transparent to the PU while only a state restoration is necessary on the PE side. A demo application using the RSerPool session layer in a Distributed Computing application is described in .

Applying RSerPool for distributed computing applications, the duties of the RSerPool architecture are still limited to the management of pools and independent sessions only. It is in particular a non-goal to provide functionalities like data synchronization among sessions, user authentication, accounting or the support for more than one administrative domain. Such functionalities are considered to be application-specific and are therefore out of the scope of RSerPool.

A proof of concept implementation of a Distributed Computing application based on the RSerPool prototype rsplib can be found at . This system provides a fractal graphics computation service; the failover procedure is handled by ASAP cookies.

The protocols used in the Reliable Server Pooling architecture only try to increase the availability of the servers in the network. RSerPool protocols do not contain any protocol mechanisms which are directly related to user message authentication, integrity and confidentiality functions. For such features, it depends on the IPSEC protocols or on Transport Layer Security (TLS) protocols for its own security and on the architecture and/or security features of its user protocols. The RSerPool architecture allows the use of different transport protocols for its application and control data exchange. These transport protocols may have mechanisms for reducing the risk of blind denial-of-service attacks and/or masquerade attacks. If such measures are required by the applications, then it is advised to check the SCTP applicability statement RFC3257 for guidance on this issue.