Distributed wisdom: analyzing distributed-system performance - latency vs. throughput

R. Friedman, E. Hadad
2006 IEEE Distributed Systems Online  
Many textbooks and articles have discussed the fact that latency and throughput aren't opposites. Consider the well-known comparison of the throughput of a modern cargo ship packed with tapes on a two-week journey with the bandwidth in today's fastest networks. The cargo ship wins big-time. Clearly, if you wish to send a small packet, the Internet is a better option. However, for transferring a very large database, low-tech options would prove faster. Technical people often forget this obvious
more » ... bservation. For example, in the summer of 1997, one of us attended a talk by a founder and the CTO of a leading search engine. He explained that his company's Web site was hosted in California with a mirror on the East Coast. His search engine updated its content twice a week, because the company wants to keep the main site and its mirror synchronized, and copying the entire database over the Internet took 72 hours. An attendee asked, "In this case, why don't you store it on a tape and send it with overnight delivery?" The speaker paused for a few seconds and replied, "Hmm, that's a good point. We never thought about it." You might be asking yourself what this anecdote has to do with modern distributed systems. The answer is that the distributed-systems research community still often ignores the fact that as long as the latency is reasonable, throughput is really what matters. In particular, research papers often break down the latency costs, but how often do we see a breakdown of the throughput-limiting factors? In particular, the parts of the latency that are purely CPU-bound affect throughput, yet other parts that involve networking activity and waiting in buffers and queues often have a much smaller impact on throughput. This is true because modern software typically is multithreaded, so when 1 IEEE Distributed Systems Online January 2006
doi:10.1109/mdso.2006.2 fatcat:eq4ovpd7lnberhkb2w7jwr5lka