|
19.
06.
2015
Multi-RAPL - Sharing RAPL energy counters between applications
Till Smejkal
TU Dresden
Vortrag zum Forschungsprojekt
"Running Average Power Limit", a feature present in current Intel processors,
provides the functionality to measure the energy consumption of a system at a
relatively high granularity. Being able to utilize and account this measured energy
to the programs running on the system enables various possibilities for embedded and
mobile as well as HPC or cloud systems, where the energy consumption of their task is
of high interest.
In this talk I will present how an adaption of the Linux scheduler allows us to use
the RAPL energy counters to measure the energy usage of any program while it is
running in an, apart from the scheduler, unchanged environment. This facilitates us
to collect accurate energy statistics for them, without the influences of other
parallel running programs, while they are executed in their normal environment. While
such measurements yield significant errors if one uses pure RAPL, Multi-RAPL can
provide good results, as it will be shown with energy consumptions for parallel
executions of the benchmarks from the NAS Parallel Benchmark suite.
19.
06.
2015
RDMA-accelerated Key-Value-stores demystified
Hannes Weisbach
TU Dresden
Vortrag zum Forschungsprojekt
Starting with Pilaf, a number of RDMA-accelerated Key-Value-stores
were published. While Pilaf features RDMA-acceleration only for
lookups, Nessie and Falafel promise to speed up write access by using
RDMA-CAS instructions. FaRM goes even further, exposing the memory of
a cluster as shared address space.
Surprisingly, HERD, using RDMA-operations only for message passing
without exposing internal data structures over RDMA, outperforms them
all.
In this talk I will give a brief introduction into the RDMA
architecture, programming with libibverbs/librdma_cm, and implementing
client-server scenarios. I will subsequently discuss design
trade-offs in Pilaf and Hydra, my own RDMA Key-Value-store
implementation, and show why HERD achieves superior performance to
both implementations.
|
