Explore techniques to handle non-monotonically increasing series

[GoogleCodeExporter]

There are four different types of performance metric series that we've 
recognized: Monotonically Increasing (MI), Monotonically Increasing Deltas 
(MID), Non-Monotonically Increasing (NMI), and Non-Monotonically Increasing 
Deltas (NMID).

MI is what we're used to, series that are additive and always increase from one 
element to the next. This MI trace might record total number of bytes 
transferred:
a 3
b 7
c 8

MID is just the deltas of some underlying MI series. This MID trace corresponds 
to the above MI trace, instead recording the number of bytes transferred per 
second (or since the last event):
a 0
b 4
c 1

NMI is a series that isn't additive but represents some basic performance 
metric, basic in that it isn't the deltas of another underlying metric (as is 
the case for MID). This NMI trace might record memory usage of the program at 
each event:
a 4.1
b 12.3
c 5.2

NMID is just the deltas of some underlying NMI series. This NMID trace is a 
version of the above NMI trace, instead recording the change in memory usage 
since the last event:
a 0
b 8.2
c -7.1

As it stands, Perfume can only handle MI series. To support other types, what 
are some different solutions we might try, either pre-processing, deeper code 
changes, or something else? Try manually running through a few example 
traces/logs to see what happens when a given strategy is applied.

Original issue reported on code.google.com by tonyohm...@gmail.com on 11 Feb 2014 at 5:59

[GoogleCodeExporter]

Issue 361 has been merged into this issue.

Original comment by tonyohm...@gmail.com on 24 Feb 2014 at 1:18

[GoogleCodeExporter]

Original comment by bestchai on 15 Jul 2014 at 6:54

• Changed state: Fixed