Cost-effective Cloud Services for HPC in the Cloud: The IaaS or The HaaS?

Transcription

1 Cost-effective Cloud Services for HPC in the Cloud: The IaaS or The HaaS? Ifeanyi P. Egwutuoha 1, Shiping Chen 2, David Levy 1, and Rafael Calvo 1 1 School of Electrical & Information Engineering, The University of Sydney, Australia {ifeanyi.egwutuoha, david.levy, rafael.calvo}@sydney.edu.au 2 Information Engineering Laboratory, CSIRO ICT Centre, Australia shiping.chen@csiro.au Abstract In the scientific research domain, traditional High Performance Computing (HPC) refers to the use of sucomputers, grid environments and/or clusters of computers to solve computation-intensive and/or data-intensive problems. The traditional HPC systems are expensive and sometimes require huge start-up investment, technical and administrative support and job queuing. With the benefits of cloud computing, cloud services such as Infrastructure as a Service (IaaS) and as a Service (HaaS), enables scientists and researchers to run their HPC in the cloud without upfront investment associated with the traditional HPC infrastructure. In this pa we analyze the computational formance and dollar cost of running HPC in the cloud when IaaS or HaaS is leased. We find that HaaS significantly reduces the cost of running HPC application in the cloud by 20% compare to IaaS without significant impact to application s formance. We also found that there is a substantial improvement in computational formance in HaaS compare to IaaS. Keywords: HPC, cloud computing, HaaS, computation-intensive, computational formance 1. Introduction In the scientific research domain, traditional High Performance Computing (HPC) refers to the use of sucomputers, grid environments and/or clusters of computers to solve computation-intensive problems. Some common uses of HPC systems include weather forecasting, aircraft crash simulations, computational fluid dynamics for aerodynamics studies and many other computation-intensive [14], [20]. Today, HPC systems also offer new opportunities in business. For example, in financial institutions HPC systems are used in real time modelling to make informed investment decisions. The most powerful HPC systems are ranked on top500 [1]. Huge capital is needed to acquire the HPC systems, this makes it difficult for research communities. Until recently, HPC systems would have been out of reach for most research communities. With the recent advancement in computing technologies, computation-intensive are not only executed in the traditional HPC systems but also in HPC system in the cloud. Cloud computing [2], [3], [8] is a revolutionary computing paradigm for storing data and running, including computation-intensive. It promises numerous benefits, which includes, no upfront investments. Cloud computing also reduces development time, staff (e.g., administrators), and hardware, resulting in better service and significant cost saving. It is expected that more computation-intensive will be deployed and run in HPC systems in the cloud [5], [3]. Furthermore, the Amazon Elastic Compute Cloud (Amazon EC2) cluster recently appeared in TOP500 list [1], which shows that there is a great future for HPC systems in the cloud. With Cloud computing pay-as-you-go pricing model, scientists and researchers can lease cloud services such as Infrastructure as a Service (IaaS) and as a Service (HaaS) for computation-intensive. These services are relinquished when not in use. This avoids the job queuing, which is a common phenomenon in traditional HPC system. The price model is also attractive when compared to traditional HPC systems that require huge investments capital, administrative issues and allocation policies. However, the cost of running HPC application on the cloud may be high if the cloud services are not well understood and the cost-effective cloud services chosen. If the dollar cost of running HPC in the cloud is high comparing to traditional HPC system, then the benefits of running computation-intensive application on the cloud may have been defected. HPC research communities are concerned about the cost and computational formance of different cloud services. In this pa we analyze the computational formance and dollar cost of running computation-intensive application in HPC systems in the cloud when IaaS and HaaS are leased. We find that the cost of executing computationintensive application when HaaS is leased is significantly lower compared to the IaaS model. We show that there is significant improvement in computational formance of the application on HaaS if the computation-intensive application is not a network intensive application. Our eximental setup uses the Message Passing Interface () implementation [4]. We provide our test results, but do not reveal the identify of the cloud providers, to avoid any head-

2 to-head comparisons. However, we do include the relevant technical details of the cloud s. In Section 2 we present the overview of Cloud services for HPC systems in the clouds, while Section 3 presents the eximental setup. and benchmark are presented in Section 4. Eximental results and Cost analysis are presented in Section 5 and 6 respectively, while Section 7 discusses related work. Finally, some conclusions are presented in Section Overview of Cloud Services for HPC Systems in the cloud With the advent of cloud computing infrastructures, cloud services providers such as Salesforce.com, Amazon [5], Rackspace, Baremetalcloud [6], Microsoft Azure, SoftLayer [7], Google, IBM offer different cloud services to cloud users. Some of these services offered are Software as a Service (SaaS), Platform as a Service (PaaS), Infrastructure as a Service (IaaS), as a Service (HaaS), Network as a Service (NaaS) and Storage as a Service (STaaS). Based on the capability provided by the cloud service provider, cloud computing services fall into four major competing categories [8], [9]. Application as a Service, Platform as a Service, Infrastructure as a Service and as a Service. Figure 1 shows the architecture of the cloud computing services. Software-as-a-Service (SaaS) Platform-as-a-Service (PaaS) (Develos implementing cloud ) Infrastructure-as-a-Service (IaaS) [(ization, Storage Network) as-a-service] programming language or software framework (e.g., Java EE) and ready to execute corresponding builds. PaaS cloud users deploy and run their software, without setting up servers and software stacks, without thinking about scalability or clustering, and often even without knowing how many computers or CPUs their application will run on. Infrastructure as a Service (IaaS) is similar to HaaS, but virtual machines are rented out instead of real hardware. IaaS cloud users have to install, configure, and maintain the virtual machines they rent and are free to choose the oating system and software stack they install in their VMs. Often IaaS users make use of a pre-installed and preconfigured VM image supplied by their provider as base installation. Users do not have root access to the hardware. A good example of cloud provider that offers IaaS for HPC is Amazon [5]. The Amazon Elastic Compute Cloud (Amazon EC2) offers cluster compute s for HPC. as a Service (HaaS); in this case, the cloud provider basically rents out bare-metal hardware (e.g., server/host and storage). Notable examples of cloud providers that offer HaaS are Baremetalcloud [6] and Soft- Layer [7]. Cloud users connect to HaaS via the Internet, install and configure (e.g., VMs) the server they leased. Cloud users choose HaaS, because it gives them full control of the server, oating system, and software/hardware stack, as well as the number of VMs they execute on it. Research communities do lease HaaS for computation-intensive and/or data-intensive and configure HPC systems according to their needs [8], [9]. Consequently, computationintensive that were traditionally run on HPC systems can now be executed in the cloud. Figure 2 shows the HaaS architecture and access level of the provider and user. as a Service (HaaS) Fig. 1: Cloud layered architecture [8] Dom0 Software as a Service (SaaS) is the highest abstraction level in the cloud. It offers cloud users ready-to-use online that are already deployed in the cloud. This layer is hidden from the users and managed by the SaaS providers. The users do not know where or how these are deployed, but simply use them. SaaS cloud can be accessed via the internet with any internet-ready device such as a laptops, smart phones, or ipads. This enables relatively dumb clients to form complex tasks, by shifting the real work, transparently to the user, into the cloud. A good example of SaaS is the commonly used Gmail ( services) provided by Google. Platform as a Service (PaaS) provides cloud users with a fully configured and managed computing platform, ready to run custom software developed by the users. Each PaaS platform is targeted to software developed in a specific Back end Drivers DomU1 DomU... Xen hyvisor DomU... [Disk, NIC, CPU, and RAM] CPU and RAM DomUn Users (eg., Scientist) Provider (e.g., baremetalcloud) Fig. 2: An example of HaaS architecture with level of involvement of key players 3. Eximental setup We setup eximental environments to evaluate the computational formance and dollar cost of running computation-intensive application on IaaS and HaaS. Our eximental setup includes two services we have leased

3 from two cloud service providers; for the purpose of avoiding head-to-head comparison of the two cloud providers, we call them Cloud-A and Cloud-B. Cloud-A offers IaaS in different kind of cluster s for HPC : for example, cluster compute s. Cloud-B offers HaaS which can be configured to run HPC. 3.1 Cluster Compute s from Cloud-A (IaaS) Cloud-A is one of the major cloud service providers. They offer IaaS in different s for HPC. Table 1 shows a sample of cluster compute s with price details of on-demand s from cloud providers. The clusters compute s are available with commonly used Oating System (OS) (Windows and Linux) in 32- bit and 64-bit platforms. For our eximents, we choose the Ref-C virtual in the Table 1 because it is widely used for HPC. The s use Xen full virtualization. The I/O network communication between the cluster is 10 Gigabit Ethernet. In order to compare the computational formance and dollar cost of running HPC when IaaS and HaaS services are leased. We leased a cluster compute with a total of 16 processors. The details of the leased cluster compute are shown in Table 2. We installed Open 1.6 [23] on the node. Open is an open source implementation of the Message Passage Interface (). Table 1: and HaaS s from Cloud-A and Cloud-B type Memory CPU Disk Ref-A Ref-B Ref-C Ref-D Ref-E Ref-F Ref-G Ref-H 30 GB 2x2.0 GHz (sixteen-core) 244 GB 2 x Intel Xeon E (eightcore) 22 GB 2 x Intel Xeon X5570 (quadcore) x Intel Xeon GB X5570 (quadcore) 96 GB 2x2.13 GHz DDR3- E5606 (eightcore) GB 2x2.66 GHz DDR3- X5650 (twelvecore) GB 2x2.0 GHz E OctoCore (sixteen-core) 32 GB 2x2.0 GHz (eightcore) Cost of for Linux 500 GB $ GB $ GB $ GB $ GB, 7200RPM 300GB, 10000RPM GB GB 1.25 Cost of Window $1.800 $3.831 $2.600 $1.610 $1.19 $0.93 $1.59 $ HPC system on HaaS in the cloud As explained in Section II, HaaS allows users to have full control of the system and control environment for measuring system formance and other available eximents. This enables users to determine the number of VMs to be deployed for HPC. We have leased an HaaS (Ref-G) with 64GB RAM from Cloud-B. Table 1 shows some of the cloud services that the HaaS providers offer that are similar to cluster compute s that Cloud- A offers. The table also gives a summary of HaaS and price of the service leased. The communication network between each HaaS is a 1 Gigabit Ethernet. The summary of the VM we provisioned on the HaaS is shown in Table 2. We installed Xen hyvisor [11] on the host. Xen hyvisor is an open source, industry standard virtualization technology. Linux Oating System (Ubuntu bit) runs on top of the Xen hyvisor. We imported our pre-configured para-virtualised guest OS (Ubuntu bit) on the HaaS. The pre-configured paravirtualised guest reduces the time to setup the HPC system on the HaaS. A para-virtualized OS uses a modified kernel, and reduces the size of the image. The VM is configured to have 16 processors with 60GB memory and 200GB hard drive. We installed Open on the node. This setup is almost equivalent to the cluster compute s we leased from Cloud-A. The setup also allow us to have a good comparison environment for IaaS and HaaS in terms of computational formance and dollar cost. Table 2 shows both the IaaS and HaaS environments we used. Table 2: Computational environment for IaaS and HaaS Cloud&A,)VM)of)IaaS Cloud&B,)VM)of)HaaS RAM:%24%GB% RAM:%60%GB% Architecture:%%%%%%%%%%x86_64 Architecture:%%%%%%%%%%x86_64 CPU%op;mode(s):%%%%%%%%32;bit,%64;bit CPU%op;mode(s):%%%%%%%%64;bit CPU(s):%%%%%%%%%%%%%%%%16 CPU(s):%%%%%%%%%%%%%%%%16 On;line%CPU(s)%list:%%%0;15 Thread(s)%%core:%%%%16 Thread(s)%%core:%%%%2 Core(s)%%socket:%%%%1 Core(s)%%socket:%%%%4 CPU%socket(s):%%%%%%%%%1 NUMA%node(s):%%%%%%%%%%1 NUMA%node(s):%%%%%%%1 Vendor%ID:%%%%%%%%%%%%%GenuineIntel Vendor%ID:%%%%%%%%%%%%%GenuineIntel CPU%family:%%%%%%%%%%%%6 CPU%family:%%%%%%%%%%%6 Model:%%%%%%%%%%%%%%%%%%%26 Model:%%%%%%%%%%%%%%%%%26 Stepping:%%%%%%%%%%%%%%%%5 Stepping:%%%%%%%%%%%%%%5 CPU%MHz:%%%%%%%%%%%%%%% CPU%MHz:%%%%%%%%%%%%%%% Hyvisor%vendor:%%%%%Xen Hyvisor%vendor:%%%%%Xen ization%type:%%%full ization%type:%%%para 4. and benchmark We used a commonly used HPC benchmark and real HPC application to analyze and evaluate the

4 running on IaaS and HaaS services. The benchmark was the High Performance Linpack (HPL) benchmark [12] and the application was ClustalW- [14]. We desribe them below. HPL [12] is a benchmark that is commonly used to evaluate the computational formance of HPC systems for example, top500 [1]. It measures the floating execution rate of linear equations based on the problem size. We executed the HPL benchmark with five different problem sizes of 2,000, 4,000, 6,000, 8,000 and 10,000 on the both cloud services on VMs from IaaS and on HaaS. The execution of each the problem sizes was carried twice and the average execution time calculated. The five different problem sizes enable us to obtain different wall clock execution times of HPL. We recorded the wall clock execution time for each problem size. We used the wall clock execution time to analyse the dollar cost and computational formance of the both platforms. Figure 3 show the results obtained on computational-formance. ClustalW- [14] is a parallel implementation of ClustalW [15] which is based on. ClustalW is a tool that is widely used in bioinformatics for multiple alignments of nucleic acid and protein sequences. It uses three alignments steps: pairwise alignment, guide-tree generation and progressive alignment. We ran a sample of A full multiple sequence alignment, A guide tree only seqence alignment, and A multiple sequence alignment out of an existing on nodes from IaaS and from HaaS. We recorded the execution time of the three alignment steps to compare time to finish executions with both IaaS and HaaS. The results are shown in figure Results and Discussion One of the major attractions to the Cloud-A cluster compute is that it is relatively easy to set up the clusters compared to setting up a cluster in HaaS. However, some level of technical knowledge is required to setup cluster on Cloud-A that will run HPC due to varying needs of HPC. In order to reduce the time to set up an HPC system on HaaS s in the cloud, we uploaded our pre-configured para-virtualized image to the cloud. There are also similar VM images which can be downloaded from different sites. We estimated that this technique reduces the set up time by up to 80%. We did not compare the time to setup HPC system in Cloud-A (IaaS) cloud and in Cloud-B (HaaS) because setup time varies with individuals technical exiences. From the computational formance result of the HPL benchmark shown in figure 3, we can see that the wall clock execution time of HPL benchmark on a provisioned on HaaS is shorter when compared to IaaS provided by Cloud-A. We achieved this because the memory of the virtual s deployed on HaaS is 60GB. We chose to allocate this amount of memory to our virtual because we can predict the memory needed. This option is not available for the IaaS (users cannot change the memory of the virtual chosen). We also have full control of the and virtual s. Wall clock execu4on 4me in seconds Cloud- A (VM of IaaS) Cloud- B (VM of HaaS) 2,000 4,000 6,000 8,000 10,000 HPL problem sizes on 16 processors Fig. 3: Computational formance of High Performance Linpack on 1 node with 16 processors As shown in Figure 3, executing the HPL on 1 node with 16 processor eliminates the bandwidth inequality on both providers. The virtual s HaaS out forms IaaS. This is because we have full control of the running of our HaaS and we allocated higher memory to VM on HaaS. On IaaS, other VM s may have been hosted on the hardware which may have affected the formance of the application running on our lease IaaS. As shown in [19], high resource allocations on infrastructure affect running on VMs. The ClustalW- results is shown in Figure 4. Cloud-A IaaS uses 10 Gigabit Ethernet network, whereas HaaS we leased uses a 1 gigabit Ethernet network. We could have benchmarks with the same bandwidths, however the two major providers of HaaS do not have 10 gigabit Ethernet network. The results in Figure 4 show that there is no significant impact on application running on IaaS and on virtual s on HaaS. 6. Cost Analysis At the time of writing, Cloud-A offers different price models to their cluster compute customers; The primary price model which is widely used is called ondemand s. The on-demand s price model allows users to pay ly without contract while other price models may require up front payments and/or contracts. Cloud-B offers their customers a pay-as-you-go price model, which is similar to on-demand prices offered by Cloud-A. Therefore we use on-demand price

5 Wall clock execu+on +me in seconds Cloud- A (VM of IaaS) Cloud- B (VM of HaaS) Dollar cost $ Cloud- A (VM of IaaS) Cloud- B (VM of HaaS) 0 A full mul+ple sequence alignment A guide tree only sequence alignment A mul+ple sequence alignment out of an exis+ng tree 0 2,000 4,000 6,000 8,000 10,000 HPL problem sizes on 16 processors Fig. 4: Performance of ClustalW- application on 16 processors to compare the cost of running computation-intensive on both cloud services. In addition to the ondemand/pay-as-you-go s prices, there are charges, which are charged for some cloud services such as network bandwidth and IP addresses which we do not consider to avoid complexity. We used the results obtained from HPL benchmarking to analyse the cost. As previously used in a similarly cost analysis [18], we assume that 1 second is equal to ly rate which the both cloud providers offer. This also allows us to do the analysis without paying the for s the eximent would have cost. We used the prices of the leased services as shown in Table 1 and 2. The cost analysis computation of IaaS and HaaS is shown in figure 5. Based on the computational formance and cost analysis it appears that it is more cost effective to lease HaaS and configure the HPC systems. Cloud service users of HaaS have full control of the hardware as well as the VMs they provisioned. Application formance and other metrics can be easily measured. From the result, it seems that the cost of running HPC can be reduced by 20% when HaaS is leased. 7. Related work Cloud computing is a revolutionary computing paradigm for storing data, running, including computation-intensive. Cloud computing promises numerous benefits, which includes no up front investments for HPC, which is attractive, compared to traditional HPC systems. Many studies have evaluated the suitability of HPC systems in the cloud and showed that it is expected that more computation-intensive HPC will be run in the cloud HPC than Fig. 5: The cost analysis VM of IaaS and VM of HaaS traditional HPC systems [16]. Furthermore, the Amazon Elastic Compute Cloud (Amazon EC2) cluster recently appeared in TOP500 list [1] in year 2010, which shows that there is a viable future for HPC systems in the cloud. Many past researches evaluating of HPC on HPC systems in the Cloud with emphasis on Amazon EC2 have been carried out. These investigations focus on the formance of Amazon EC2 and Traditional HPC systems [16], [17], [18], [19]. Carlyle et al. [17] studied the cost effective HPC System. They show that it is cost effective for institutions like Purdue University to oate a community/traditional cluster than to lease HPC resources from Amazon EC2. This study clearly shows that Amazon on-demand cluster compute s prices are not cost effective for HPC for some institutions. Their work focuses on Amazon EC2 service IaaS and traditional HPC systems. Deelman et al. [18] in their work on The Cost of Doing Science on the Cloud: The Montage examples ; show that the cost of cloud services could be significantly reduced without significant impact on application formance, if the right storage and compute resources are provisioned. However, they did not consider different platforms like HaaS. We extended their work, demonstrating that HaaS can significantly reduce the cost of running computationintensive application on HPC in the cloud. Ekanayake and Fox [19] compare HPC with different needs and showed the formance of with latency. However, they did not compare the cost of executing computation-intensive application on different services such as IaaS and HaaS. Yao et al. [21] showed that optimal cost-formance ratio can be achieved with th appropriate cloud. However, they did not consider cost and computational formance

6 when IaaS and HaaS are leased. To the best of our knowledge, our work is different from other work in that we study the computational formance and dollar cost of running computation-intensive application in HPC in the cloud when IaaS and HaaS are leased. We eximentally show that the dollar cost of running computation-intensive application can be reduced as much as 20% with HaaS without significant impact to formance. 8. Conclusions and Future Work Due to the huge capital investment required to own a traditional HPC systems which typically involves job queuing, using an HPC system in the cloud is a good alternative. Cloud computing offers IaaS and HaaS for deployment of cluster s, which can be used to run computation-intensive. IaaS provides almost ready to use clusters with minimal deployment installation tasks. With HaaS, virtual machines can be provisioned to run computation-intensive application. We have conducted eximental analysis to determine the formance and cost when cloud services IaaS and HaaS are leased to run computation-intensive application. We showed that the dollar cost of running computation-intensive application in the cloud can be reduced by as much as 20% when HaaS is leased. We showed that there is no significant impact in formance of the when executed on the leased HaaS. Acknowledgment The authors would like to thank Bran Selic for providing valuable comments and suggestions. References [12] Petitet, A., Whaley, C., Dongarra, J., and Cleary, A., (2008, Sept), HPL Benchmark, [Online], benchmark/hpl/ [13] Egwutuoha, I. P., Levy, D., Selic, B., and Chen, S., A survey of fault tolerance mechanisms and checkpoint/restart implementations for high formance computing systems, The Journal of Sucomputing, Feb 2013, /s [14] Li, Kuo-Bin. ClustalW-: ClustalW analysis using distributed and parallel computing, Bioinformatics 19, no. 12 (2003): pp [15] Thompson, Julie D., Desmond G. Higgins, and Toby J. Gibson. CLUSTAL W: improving the sensitivity of progressive multiple sequence alignment through sequence weighting, position-specific gap penalties and weight matrix choice Nucleic acids research 22, no. 22 (1994): [16] Evangelinos, C., and Hill, C. N., Cloud Computing for parallel Scientific HPC Applications: Feasibility of running Coupled Atmosphere- Ocean Climate Models on Amazon s EC2, ratio 2, no (2008): pp [17] Carlyle, A. G., Stephen L. H., and Preston M. S., Cost-effective HPC: The community or the Cloud?, In Cloud Computing Technology and Science (CloudCom), 2010 IEEE Second International Conference on, pp IEEE, 2010 [18] Deelman, E., Singh, G., Livny, M., Berriman, B., and Good, J., The cost of doing science on the cloud: the montage example, In Proceedings of the 2008 ACM/IEEE conference on Sucomputing, (2008, November), (p. 50), IEEE Press. [19] Ekanayake, J., and Fox, G., High formance parallel computing with clouds and cloud technologies Cloud Computing (2010): [20] Fox, G. C., and Coddington, P. D., An overview of high formance computing for the physical sciences, In Proceedings of Mardi Gras Conference: High Performance Computing and Its Applications in the Physical Sciences, [21] Yao, J., Ng, A., Chen, S. et al, A Performance Evaluation of Public Cloud Using TPC-C Benchmark, The 1st International Workshop on Analytics Services on the Cloud (ASC 2012), in conjunction with ICSOC 2012: 1-7. [22] Penguin Computing. services/hpc-cloud/pod/architecture. [23] Open, [1] [2] Armbrust, M., Fox, A., et al, A view of cloud computing, Communications of the ACM, 53(4), pp , [3] Mell, P., and Grance, T. The NIST definition of cloud computing (draft), NIST special publication, 800, (2011), pp [4] Message Passing Interface Forum, : A message-passing interface standard, International Journal of Sucomputer Applications, 8(3/4): , [5] Amazon. (2013), [Online], [6] Baremetalcloud [Online], index.php/en/ [7] SoftLayer (2013), [8] Rimal, B. P., Choi, E., and Lumb, I., A taxonomy and survey of cloud computing systems, In INC, IMS and IDC, 2009, NCM 09, Fifth International Joint Conference on (pp ). IEEE, [9] Egwutuoha, I. P., Chen, S., Levy, D., Selic, B. and Calvo, R., A Proactive Fault Tolerance Approach to High Performance Computing (HPC) in the Cloud, in The 2nd International Conference on Cloud and Green Computing, Xiangtan, Hunan, China, 2012, pp [10] [11] Xen hyvisor, [Online],