\reviewtitle{METRIC: a kernel instrumentation system for distributed environments}
\reviewlabel{mcdaniel77metric}
\reviewauthor{Gene McDaniel}

Developed in 1977 at Xerox PARC, METRIC runs probes on each
workstation connected via an Ethernet.  Each probe sends events to one
or more accountants. The accountants may keep all events, drop
events, or selectively keep information (was this explained?).  Using
the information at the accountant, the analyst is able to answer
questions along the lines of ``What is the average time to access
files in the system?''  METRIC is not able to --- nor do the authors mention
--- trying to answer inter-machine dependency questions.  METRIC was
used by 15 machines at once.

Events are not counters, but mini-snapshots of when events begin or
end (or begin and end?).  They provide a back-of-the-envelope
calculation that 80 machines could send 75 events per second (at 14
bytes per event) without ``seriously clogging the Ethernet.''  The
system work with a loss rate of about $.5\%$ of events.  Events have
types and subtypes (a good idea and one used in the next paper,
Yaghmour).  

Q1. What happens if events overlap coming from one workstation?  Can
this happen?

Q2. What would be the point in the ``network'' configuration?  What
do multiple accountants provide?
A2. Maybe could hook up multiple Ethernets to each machine to get
more logging capacity.