24.04.2014 Views

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

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

e-­‐FISCAL: www.efiscal.eu <br />

EC Contract Number: 283449 <br />

5.3 e-FISCAL sample <strong>and</strong> main findings<br />

e-­‐FISCAL received 28 high quality answers from 16 countries which constitute a diverge sample from different <br />

types of centres in terms of size (small <strong>and</strong> large), in terms of nature (HPC <strong>and</strong> HTC), with a low or high number of <br />

staff, etc. e-­‐FISCAL was able to calculate several metrics, including <strong>cost</strong> per core hour, CAPEX/OPEX ratio, <br />

depreciation rates, <strong>cost</strong> distribution, electricity PUE. The main findings can be summarised as follows: <br />

• Cost effectiveness: In-­‐house HPC/HTC e-­‐<strong>Infrastructure</strong>s are <strong>cost</strong>-­‐effective with the high average utilisation <br />

rates (65-­‐75%) <strong>and</strong> depreciation rates (5 years) recorded. However, per application <strong>cost</strong> analysis is needed <br />

<strong>and</strong> e-­‐FISCAL did not have the resources to execute several comparative application case studies. <br />

• Cost structure: OPEX dominates the <strong>cost</strong>s (around 70%) – Personnel <strong>cost</strong>s are around half of the total <strong>cost</strong>s, <br />

while personnel <strong>cost</strong>s vary a lot around Europe. <br />

• Comparison with state of the art: The e-­‐FISCAL findings are in-­‐line with the results of similar studies identified <br />

in the state of the art through literature review. <br />

• HPC-­‐ HTC comparison: It appears that the HPC sites of the e-­‐FISCAL sample (national Tier-­‐1 centres) are in <br />

average cheaper than corresponding HTC ones. Possible reasons for this are the higher number of cores per <br />

CPU, the state of the art energy systems, <strong>and</strong> the fact that larger sites have in most cases less FTEs per core. <br />

• Cost on the basis of size (no of cores) -­‐ Big vs. small sites: It appears that sites with 1000 cores. Indicative values for <strong>cost</strong> per core hour are around € 0.09 for the first <br />

category <strong>and</strong> € 0.05 for the latter. <br />

• Cost structure heterogeneity: there is heterogeneity in the <strong>cost</strong> structure of sites/centres (in terms of <strong>cost</strong> <br />

categories). However, it is rather evident that in the vast majority of cases OPEX is not less than 40% of total <br />

<strong>cost</strong>s <strong>and</strong> in the average case OPEX dominates <strong>cost</strong>s. <br />

• In-­‐house HTC <strong>and</strong> HPC vs. commercial Clouds; Benchmarking: Cloud service providers usually have a wide <br />

portfolio of instances <strong>and</strong> services. <br />

o Comparisons between in-­‐house <strong>and</strong> commercial clouds have to be made very carefully, trying to <br />

compare like with like. Benchmarking is thus very relevant in this exercise as it identifies comparable <br />

performances between in-­‐house configurations <strong>and</strong> cloud instances, especially because the cloud <br />

instance description might not always be easily comparable. <br />

o Performance degradation of cloud virtual machines with the in-­‐house e-­‐<strong>Infrastructure</strong> have to be <br />

accounted for through appropriate weight factors, especially in cases where virtual machines are <br />

compared to bare metal. <br />

o With the small-­‐scale benchmarking exercise a 27% <strong>and</strong> 43% degradation has been averaged for HTC <br />

<strong>and</strong> HPC respectively. <br />

o Taking these degradations into account <strong>and</strong> comparing Amazon EC2 with the e-­‐FISCAL HTC <strong>and</strong> HPC <br />

<strong>cost</strong>s, it is obvious that: <br />

Amazon st<strong>and</strong>ard reserved instance prices are comparable with e-­‐FISCAL ones <strong>and</strong> sometimes <br />

cheaper (the 3-­‐year reserved ones) <br />

Amazon st<strong>and</strong>ard On-­‐dem<strong>and</strong> ones are always more expensive than e-­‐FISCAL <br />

Amazon cluster compute ones (HPC in the cloud) are always much more expensive (1.5 to <br />

around 5 times) <br />

• Several disclaimers have to be taken into account though as analysed in the previous <br />

section such as comparing 2012-­‐2013 prices with 2011 <strong>cost</strong>s, the exclusion of <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 />

71

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!