X-Trace: A Pervasive Network Tracing Framework

R. Fonseca, G. Porter, R. H. Katz, S. Shenker, I. Stoica, "X-Trace: A Pervasive Network Tracing Framework," NSDI'07, (April 2007).

One line summary: This paper describes a diagnostic tracing framework for distributed or networked applications called X-Trace that is intended to be used as a debugging tool.

Summary

This paper describes a tracing framework for distributed or networked applications called X-Trace. X-Trace is intended to be used as a diagnostic tool. It works by tagging each application task with metadata including an identifier, and each network operation originating from that task is tagged with that task identifier, along with data relevant to that operation. From a trace of these tags, it is possible to construct a task tree, which can be used to debug networked applications. In X-Trace, the trace request is sent in band, while the collected trace data itself is sent out of band back to the requesting source. The authors implemented X-Trace and tested the performance of an web server and database back-end with X-Trace enabled. They observed a small decrease in performance, but feel that this is justified given X-Trace is meant to be used for debugging. The authors also provide a number of usage scenarios.

Critique

I didn’t think that this paper was a particularly great idea. It doesn’t seem that difficult of a concept to come up with, but it does seem very unlikely that it would actually become widely used because it requires modifications to applications and network protocols. It also would not be very useful if partially deployed, as the authors mention. Other drawbacks that the authors mention include security concerns, non-tree request structures that can’t be captured, the difficulty in distinguishing lost reports from other problems and the difficulty in constructing causal connections if reports are lost, the extra traffic caused by X-Trace, and the difficulty in ascertaining whether X-Trace is actually a useful tool in helping system administrators and other users. The authors treatment of these issues did not really convince me that X-Trace was still a worthwhile consideration. I am not altogether convinced that alternative and potentially currently available diagnostic tools and methods could be just as, if not more, useful, especially considering the potential difficulty in deploying X-Trace. I think something that requires no modifications to existing applications and protocols is vastly more appropriate.