Skip to content
Snippets Groups Projects
energy.md 2.27 KiB
Newer Older
Branislav Jansik's avatar
Branislav Jansik committed
# Energy Saving

Branislav Jansik's avatar
Branislav Jansik committed
Due to high energy prices and reductions in funding, we implement a set of energy saving measures at the  IT4Innovations National Supercomputer Center. The measures are selected such as to minimize the performance impact and achieve significant cost, enegy and carbon footprint reduction effect.
Branislav Jansik's avatar
Branislav Jansik committed

Branislav Jansik's avatar
Branislav Jansik committed
The energy saving measures are effective as of 31.1.2023.
Branislav Jansik's avatar
Branislav Jansik committed

## Karolina

### Measures

Branislav Jansik's avatar
Branislav Jansik committed
Core frequency capping is implemented for the Karolina supercomputer:
Branislav Jansik's avatar
Branislav Jansik committed
|Measure                                                  | Value   |
|---------------------------------------------------------|---------|
Branislav Jansik's avatar
Branislav Jansik committed
|Compute nodes **cn[001-720]**<br> CPU core frequency [capping][3] | 2.100 GHz |
Branislav Jansik's avatar
Branislav Jansik committed
|Accelerated compute nodes **acn[001-72]**<br> CPU core frequency capping | 2.600 GHz  |
|Accelerated compute nodes **acn[001-72]**<br> GPU core frequency capping | 1.290 GHz |
Branislav Jansik's avatar
Branislav Jansik committed

### Performance Impact

Branislav Jansik's avatar
Branislav Jansik committed
The performance impact depends on the [arithmetic intensity][1] of the job.
The [arithmetic intensity][2] is a measure of floating-point operations (FLOPs) performed by a given code (or code section) relative to the amount of memory accesses (Bytes) that are required to support those operations. It is defined as a FLOP per Byte ratio (F/B). It is a characteristics of the particular computational algorithm. 

In general, the processor frequency capping has low performance impact for computation with intensity below
the [ridge point][2], such as memory bound computations. For intensive, CPU bound computations, the impact is directly
proportional to the frequency reduction.

On Karolina, relative time increase factor up to 1.3 is observed for  intensive workloads on CPU, up to 1.1 on GPU. No slowdown is observed for memory bound workloads.
Branislav Jansik's avatar
Branislav Jansik committed

### Energy Saved

Branislav Jansik's avatar
Branislav Jansik committed
The enegy efficiency in floating point operations per joule is increased by about 30% for CPU workloads, about 25% for GPU workloads. The efficiency depends on the arithmetic intensity, however energy savings are always achieved.
Branislav Jansik's avatar
Branislav Jansik committed
## Barbora

None implemented yet.

## NVIDIA DGX-2

None implemented yet.

Branislav Jansik's avatar
Branislav Jansik committed
## Complementary Systems 
Branislav Jansik's avatar
Branislav Jansik committed

None implemented yet.

Branislav Jansik's avatar
Branislav Jansik committed
[1]: https://en.wikipedia.org/wiki/Roofline_model
Branislav Jansik's avatar
Branislav Jansik committed
[2]: https://people.eecs.berkeley.edu/~kubitron/cs252/handouts/papers/RooflineVyNoYellow.pdf
[3]: https://slovnik.seznam.cz/preklad/anglicky_cesky/capping