Lect13 notes - Efficient Multicore Global Power Management
Problem being solved
- Hard limit to temperature distribution
- Has long time constants (ms range)
- Hard limit to power distribution
- Shorter time constant
- Can cause voltage drops when stressed
- Maximize performance for a given power budget
- Never exceed power budget, also save energy
- Minimize power without impacting performance much
- How to use (HW only) dynamic global control
Users
- GPP
- Multicore (high core count)
- Multiprogramming only - no improvement/degradation for any specific application
- Ignore multithreaded apps because of synchronization
What is unique
- Dynamic, per core control
- Hardware controller
- Hierarchical control
- Local for emergency situations
- Global for optimizing impact
- Straightforward engineering/unique evaluation
- New instruments
- Experimental infrastructure
- Multicore simulator with power and fast evaluation mode
Evaluation
What they did
- Multiprogrammed with different combinations of SPEC2000 applications
- Chose specific applications to exhibit desired behaviour
- Power/performance tradeoffs for different schemes
- Assume hard limit of x% peak power
- Compare MaxBIPS to 3-mode global DVFS and optimistic static VFS
- Use an oracle for per-core VFS
- Changed the number of cores
- Adapt to budget change
Critiques
- Didn’t consider cost of per-core DVFS
- Power overshoots?
- How often? How high?
- Should change the oracle explore times and other oracle aspects
- Oracle may not be the upper bound
- Oracle needs more information than test subject to really be an oracle
- Doesn’t match goal of demonstrating maximum power without beating power budget
- Don’t consider power budget/overshoots
- Bad to use the same model for heuristic and evaluation (self-fulfilling prophecy)
- Not a huge flaw, but unconvincing
Other notes
Why power and temperature?
- Directly related
- Controlling both is important - can lead to problems
- Thermal runaway
- Chip lifetime decreases
- Lose cooling capacity over time
Linear DVFS model
- Assumes 10% voltage change translates to 10% frequency change
- Very optimistic
- More accurate model likely 10% V → 30–40% freq
- Not unique about this paper
DVFS alternatives
- Fetch gating
- Can’t be better than linear
- Small gain without voltage change
