1 HPC as Service and Scientific Applications Bruno Schulze
2 Clouds in Support of Scientific Applications Several studies have been carried out to check possibilities and limitations of clouds in providing support to scientific applications
3 The UberCloud Experiment Working teams having the opportunity to experience HCP resources in their research... Analyzing the accessing of HPC resources... Allowing about 1000 cores per hour...
4 Cloud Challenges (2013)
5 Application Areas
6 clouds emerging to provide additional resources, minimizing cost in the acquisition of new resources Clouds as Additional Resources
7 Virtualization technologies, the basic elements of clouds' infrastructure, and despite of significant advances still presenting limitations when confronted to the demands for high computational power and communication, by several scientific applications Clouds as Additional Resources
8 an improved usage of virtualized resources demands a better / deeper understanding of the characteristics of cloud architectures being used... Clouds as Additional Resources
9 interoperable and standard-based cloud federations able to achieve scientific applications requirements Clouds as Additional Resources
10 The HCP Cloud from Bull Shared Public all the simplicity of pay-per-use Reserved Public garanteed resources at optimum cost On-site Private your infrastucture, Bull's expertise Hosted Private: Innovative completely delagated /
11 Microsoft Azure Offering A8-type instances with 8 cores / 56 GB mem at $2,45/h and A9-type instances with 16 cores / 112 GB at $4,90/h... Allowing 40Gbits/s InfiniBand interconnections with remote direct memory access (RDMA), and for high performance MPI based clusters … InfiniBand interconnections only available under Windows server environment...
12 Amazon Elastic Cloud Computing API NAMEMem.(GB)Cores/Compute UnitsInstance Storage (GB)32/64 bitsI/O PerformanceEBS OptimizableCost (linux/h) t2.micro11/Variable0 (EBS only)32/64LowNo$0.013 t2.small21/Variable0 (EBS only)32/64LowNo$0.026 t2.medium42/Variable0 (EBS only)32/64LowNo$0.052 m1.small1.71/116032/64ModerateNo$0.060 m1.medium3.751/241032/64ModerateNo$0.120 m1.large7.52/485064Moderate500 Mbit/s$0.240 m1.xlarge154/ High1000 Mbit/s$0.480 m3.xlarge154/130 (EBS only)64Moderate500 Mbit/s$0.500 m3.2xlarge308/260 (EBS only)64High1000 Mbit/s$1.000 m2.xlarge17.12/ ModerateNo$0.410 m2.2xlarge34.24/ High500 Mbit/s$0.820 m2.4xlarge68.48/ High1000 Mbit/s$1.640 c1.medium1.72/535032/64ModerateNo$0.145 c1.xlarge78/ High1000 Mbit/s$0.580 cc1.4xlarge232/ Xeon X Very High (10 Gbit/s)?$1.300 cc2.8xlarge60.52/88 2 Xeon E Very High (10 Gbit/s)Not necessary$2.400 cr1.8xlarge2442/88 2 Xeon E (SSD)64Very High (10 Gbit/s)Not necessary$3.500 cg1.4xlarge22 2/ Xeon X Nvidia Tesla M2050 GPU Very High (10 Gbit/s)Not necessary$2.100 hi1.4xlarge60.516/35 8 cores + 8 hypert2*1024 (SSD)64Very High (10 Gbit/s)Not necessary$3.100 hs1.8xlarge11716/35 8 cores + 8 hypert48000 (24*2TB)64Very High (10 Gbit/s)Not necessary$
13 Google Compute Engine Machine Name Machine Type Memory (GB) Virtual Cores On-Demand Cost / Hour n1-standard-1Standard3.751$0.070 n1-standard-2Standard7.52$0.140 n1-standard-4Standard154$0.280 n1-standard-8Standard308$0.560 n1-standard-16Standard6016$1.120 n1-highmem-2High Memory132$0.164 n1-highmem-4High Memory264$0.328 n1-highmem-8High Memory528$0.656 n1-highmem-16High Memory10416$1.312 n1-highcpu-2High CPU1.802$0.088 n1-highcpu-8High CPU3.604$0.176 n1-highcpu-8High CPU7.208$0.352 n1-highcpu-16High CPU $0.704 f1-microShared Core0.601$0.013 g1-smallShared Core1.701$0.035
14 Dwarfs in application classification Cell color indicating presence of that computational pattern in that application: red/high; orange/moderate; green/low - The Berkeley Par Lab: Progress in the Parallel Computing Landscape
15 Energy X Performance... Power usage effectiveness... New processors, GPUs, CPUs, ARM based … Low latency interconnections... Actual cases about HPC public (and hybrid) clouds... Project assessment about using cloud solution... Careful analysis on cost-effectiveness... Successful applications... Remarks
16 Missing Middle
17 Performance x Cost
18 Thanks Bruno Schulze
19 Using clouds in support of scientific applications has different inherited characteristics than commercial ones.. Clouds as Additional Resources