Pradyumna Kaushik
|
41206dd82e
|
refactored name of CapThreshold to LowerCapLimit. Added comment to mention that floating point operations can lead to precision loss.
|
2017-03-09 19:20:13 -05:00 |
|
Pradyumna Kaushik
|
87bd8d7cf0
|
Added logging when there is no victim left to cap.
|
2017-03-09 19:17:06 -05:00 |
|
Pradyumna Kaushik
|
2939943afd
|
Made sure that the victim that is capped the most is the one picked to be uncapped. This is to reduce the chances of benchmark starvation. Refined comments and logging.
|
2017-02-22 20:09:04 -05:00 |
|
Pradyumna Kaushik
|
373ba63933
|
fixed bugs. Made sure that we don't cap a victim below the threshold. Made sure the victims the can be capped and uncapped are maintained in both cappedVictims and orderCappedVictims.
|
2017-02-21 21:05:47 -05:00 |
|
Pradyumna Kaushik
|
726c0555ed
|
Fixed corner cases in progressive extrema -- When a node is capped and the new cap value is above a threshold then that node can be capped or uncapped in the next cycle. If the new cap value is equal to the threshold then the node cannot be capped further and can only be uncapped. When the node is uncapped and the newUncapValue is below 100 then the node can be capped or uncapped in the next cycle. If the newUncapValue is 100 then the node can only be capped.
|
2017-02-20 20:55:06 -05:00 |
|
Pradyumna Kaushik
|
d42b7a3a3b
|
changed the type of percentage in rapl.Cap(...) from int to float64. Retrofitted power-capping strategies to cap using a float64 value instead of an int. Moved common functions in loganddynamiccap.go and logAndProgressiveExtrema.go into pcp/utils.go. New power-capping strategy that builds on top of extrema, where it caps the victims at different until it can't cap further, in which case it starts uncapping them in the reverse order of capping.
|
2017-02-15 19:22:56 -05:00 |
|