EXPANSIONS for PHASE RESEARCH:
- Runtime Phase Prediction and Phase-Driven Adaptations for multiple threads:
-----------------------------------------------------------------
- Need per thread virtual counters/ GPHT states - probably within PCB
- would be much more portable if can be done with a LKM
- Key challenge: tracking fixed instruction granularities, providing PMI
interrupt for each thread simultaneously
- Runtime Phase Prediction and Phase-Driven Adaptations for multicoure/multichip:
-----------------------------------------------------------------
- very interesting issues, AMD Barcelona would be a good test bed for CMP
- Woodcrest (DP-CMP2) for multichip. Each package can be configured separately but not each
core
- how can scheduling help in such a case for "phase aligining"?
- we had already done simultaneous tracking with set_cpu_affinity, but haven't transferred to
delta_instruction tracking
- Application Phase Classification methods Exploration (Power and Performance):
-----------------------------------------------------------------
Application -> sample -> Similarity criterion -> Dimension Transformation -> Clustering -> comparison
- Incorporating MDA in runtime phase clustering:
-----------------------------------------------------------------
- PCA is good for representation; MDA is good for discrimination.
- Need to empriicially get the discriminant vectors
- Or application defines them
- Some kind of recurrence detection:
-- transitions?
-- lossy string matching?
-- simple table?
-- fuzzy clusters?
- Implementing the transition approach in HW with a management application
-----------------------------------------------------------------
--- HW reconf
--- throttle
--- CMP manage
--- power gate
- Dynamically Tuning virtual memory:
-----------------------------------------------------------------
- We had some initial attempt
- Need more insight to Vmem, CS support
- we can use our platform for power, etc
- More on Duration Prediction:
-----------------------------------------------------------------
- prediction tables and probabilistic approach
- dynamic mgmt endgoal
- Regression/Estimation/Numerical Methods for Power/Thermal Estimation:
-----------------------------------------------------------------
- Back to power estimation with better validation support,
- more detailed power estimations with counters.
- Instead of top-down, try bottom-up with power values/thermal map
- emphasis on estimators (multiple linear regressions, piecewise linear estimation, mutual info)
- Alternatively: runtime, self updating model, like neural nets (ripples will be problem in such case)
, so it updates wrt runtime power verification
- possible new approaches to phase detection; based on prior discussions:
-----------------------------------------------------------------
- shift/rotation invariant wavelet transform (wave-menu)
- windowed xcorr
- MDA, SVD, PCA, ICA, Registration, waveform matching
- statistics (time series)/ mutual info
- signal/image retrieval / feature extraction / humming recogn.
- point matching
- pattern recognition
- lossy string matching
- context free grammar (seiquitir)
- runlength/markov chains
- fuzzy clustering
- How does general phase approach compare to odour matching problem?
(compare to our neural network research)
- Leakage Mgmt/Power Gating Application with Phases:
-----------------------------------------------------------------
- Can we use phase info to guide application of where to power gate?
- What are the break-even cycles? How does it relate to phase detection?
- BBVs and The effect of Sampling:
-----------------------------------------------------------------
- Power & IPC phases with BBVs.
What if:
- we sampled every PC?
- we sampled every 10PC?
- we sampled every 1000PC?
What if?
- we didn't hash?
- hash to 1024?
- hash to 512?
- hash to 64?
- hash to 8?
- Applying the BBV recurrence info + PMC phases to thermal mgmt:
-----------------------------------------------------------------
- Use our BBV and PWR with thermal modeling, with multiple contexts
- Based on observed/predicted phases, do different context switch scenarios to better manage temperature.
- Need to consider scheduling overheads/default nature of preemptive
NEW RESEARCH DIRECTIONS:
virtualization/scalable enterprise
many core/uncore/accelerators
three-tier approach
- expose parallelism/migrate
- global power/perf/thermal
- local actions
Embedded systems/MPSoC/ASIPs
BIG phase classifier/actioner for single core
Real-time systems/SLAs for GPHT based mgmt
variability(hetero+process)
- Variation and defect tolerance
- self healing/fault tolerant
threads/TM
security
ultra low power/ultra mobile/UMPC
disposable computing
biomedical/uarch
parallelism/data-prog/graphix-throughput processing convergence
QoS caches/cores
3D arch / embedded DRAM
|