Implementation and Performance of a Communication Facility for Distributed Transaction Processing
Enrique Mafla, Bharat Bhargava · Purdue e-Pubs (Purdue University System) · 1990
We identify the problems in general purpose interprocess communication mecllanisms available for the Raid distributed database transaction processing system by conducting a series of experiments.These mechanism are CPU intensive, optimized only for remote communication and do not support multicasting.We develop a transactionoriented communication facility to address these problems.Its performance is limited only by the network device driver and the system call mechanism overheads.Sending a lOO-byte message (monocast or multicast) takes 650 microseconds, whlch includes the overheads.This is approximately 30% of the cost of the corresponding Unix commu~ IDcation facility.This communication facility demonstrates the feasibility of address spaces for structuring complex distributed transaction processing systems.The new communication facility employs shared-memory ports, a simple naming scheme, and a transaction-oriented multicasting mechanism.Local and remote communication is through ports.Ports can be accessed directly by the kernel and by user-level processes.The naming scheme used for the application and network levels avoids the use of name-resolution protocols by directly mappings the application-level name space to the network name space.Physical.multicasting is used and the need for special.pro• tocols to agree on a group address is avoided.Each transaction defines a multicasting group consisting of the set of sites involved.A group's multicasting address is a function of the corresponding transaction identifier and can be independently determined by each member of the group.The new communication facility reduces kernel overhead during transaction processing in Raid by up to 70%.