Conference · 2006
PEGASUS: Competitive load balancing using inetd
George Oikonomou, Vassilios Karakoidas, Theodore Apostolopoulos
A load balancer built on inetd in which the servers compete for work: each advertises its spare capacity to the balancer as tickets, and the balancer simply hands out the next ticket in the queue.
- Published in
- Proceedings of the 5th International System Administration and Network Engineering Conference SANE 06, 2006
- Citations
- 0 on Google Scholar, read 5 September 2026 — 30 of 30 by count
- Cite as
- OKA06
The idea
Conventional load balancing puts the knowledge in the middle: a delegator that tracks the capacity of the whole cluster and schedules accordingly. PEGASUS inverts it. Each service provider monitors its own workload and competes for requests by sending the balancer tickets — availability notifications carrying its service capacity. The balancer keeps them in a FIFO queue and, on each request, pops one and redirects. Scheduling becomes a consequence of who asked for work, and the balancer holds no global state.
Contributions
- A competitive scheduling algorithm for load balancing, with capacity advertised by the servers rather than tracked centrally.
- An implementation built on the inetd superserver, so the infrastructure is easy to configure and requires no modified TCP/IP stack — unlike the Linux Virtual Server approaches it is compared against.
- A small application-layer protocol over TCP with three messages, expressed as URLs: ticket registration, next available service provider, and service provider failure.
- A comparison with the load balancing frameworks in production use at the time, round robin DNS and LVS among them.
Written from the paper itself — the PDF linked above, which this site hosts.