D2 3 Computing e-Infrastructure cost calculations and business _models_vam1-final
D2 3 Computing e-Infrastructure cost calculations and business _models_vam1-final
D2 3 Computing e-Infrastructure cost calculations and business _models_vam1-final
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
e-‐FISCAL: www.efiscal.eu <br />
EC Contract Number: 283449 <br />
In order to calculate the performance adjusted prices for comparing the in-‐house HTC computing <strong>cost</strong>s <strong>and</strong> Cloud <br />
offerings these performance variations (for both best <strong>and</strong> worst case) are used. <br />
HPC vs. Cloud benchmarking <br />
The results of the benchmarking exercise that refer to the comparison of the in-‐house HPC instances (i.e. Stokes <br />
HPC system) against the Amazon EC2 HPC instances were included in <strong>D2</strong>.2. The following Table 12 provides a <br />
comparison of the benchmark results for both in-‐house <strong>and</strong> EC2 HPC instances using the NAS Parallel Benchmark <br />
(NPB) MPI <strong>and</strong> OMP variants. <br />
Benchmark <br />
EC2 average performance <br />
degradation VS. HPC in % <br />
EC2 range of <br />
performance variations <br />
VS. HPC in % <br />
NPB MPI – Class B -‐ 48.42 -‐ 1.02 to -‐ 67.76 <br />
NPB OMP – Class B -‐ 37.26 -‐ 16.18 to -‐58.93 <br />
Table 12: EC2 HPC VS. Stokes HPC instance @ ICHEC<br />
The MPI <strong>and</strong> OMP variants of NPB are chosen to address the most common parallelism techniques used within <br />
the HPC community. Although, medium sized problem (i.e. Class B) was chosen for various NPB kernels (i.e. BT, <br />
CG, EP, FT, IS, LU, MG <strong>and</strong> SP), the individual kernels significantly vary in size <strong>and</strong> type to cover various aspects of a <br />
real-‐world parallel/HPC application. One may argue that Class B would not be a real representative in terms of <br />
scale for a typical HPC/parallel application. But, the fact of the matter is that our objective <strong>and</strong> scope (as <br />
highlighted in the methodology section) is to identify the performance variations across EC2 <strong>and</strong> in-‐house HPC <br />
infrastructures. Of course, benchmarking a typical large-‐scale parallel/HPC application would yield more accurate <br />
results but would also require significant resources. Nevertheless, the performance variations (in Table 12) are <br />
quite evident even with Class B problem. In order to calculate the performance adjusted prices for comparing the <br />
in-‐house HTC computing <strong>cost</strong>s <strong>and</strong> Cloud offerings these performance variations are used. <br />
Therefore, the outcomes of the benchmarking exercise were used in an attempt to compare like with like services. <br />
More specifically, the loss in performance has been embedded in the <strong>calculations</strong> as follows: Price x (1+ % loss in <br />
performance). For example, the price for the on dem<strong>and</strong> instances per core hour for M/L/XL st<strong>and</strong>ard instances is <br />
€0.064/ core hour. By taking into account a performance degradation of 27% (as evidenced in the benchmarking <br />
exercise) the performance-‐adjusted price goes to € 0.081 [(0.064 + (1+0.27)) = 0.081] <br />
By comparing the <strong>cost</strong> per core hour for 2011 streaming from e-‐FISCAL <strong>calculations</strong> (Basic Case) <strong>and</strong> the Amazon <br />
performance adjusted prices for St<strong>and</strong>ard M/L/XL instances -‐ Linux per core hour we get the following results. <br />
The performance degradation used to adjust the prices of St<strong>and</strong>ard M/L/XL instances is 27% <br />
e-‐FISCAL : Financial Study for Sustainable <strong>Computing</strong> e-‐<strong>Infrastructure</strong>s <br />
Deliverable <strong>D2</strong>.3 – <strong>Computing</strong> e-‐<strong>Infrastructure</strong>s <strong>cost</strong> estimation <strong>and</strong> analysis – Pricing <strong>and</strong> <br />
Business <strong>models</strong> <br />
58