Interview Markus Schmidberger ,Cloudnumbers.com

Here is an interview with Markus Schmidberger, Senior Community Manager for cloudnumbers.com. Cloudnumbers.com is the exciting new cloud startup for scientific computing. It basically enables transition to a R and other platforms in the cloud and makes it very easy and secure from the traditional desktop/server model of operation.

Ajay- Describe the startup story for setting up Cloudnumbers.com

Markus- In 2010 the company founders Erik Muttersbach (TU München), Markus Fensterer (TU München) and Moritz v. Petersdorff-Campen (WHU Vallendar) started with the development of the cloud computing environment. Continue reading “Interview Markus Schmidberger ,Cloudnumbers.com”

High Performance Analytics

Marry Big Data Analytics to High Performance Computing, and you get the buzzword of this season- High Performance Analytics.

It basically consists of Parallelized code to run in parallel on custom hardware, in -database analytics for speed, and cloud computing /high performance computing environments. On an operational level, it consists of software (as in analytics) partnering with software (as in databases, Map reduce, Hadoop) plus some hardware (HP or IBM mostly). It is considered a high margin , highly profitable, business with small number of deals compared to say desktop licenses.

As per HPC Wire- which is a great tool/newsletter to keep updated on HPC , SAS Institute has been busy on this front partnering with EMC Greenplum and TeraData (who also acquired  SAS Partner AsterData to gain a much needed foot in the MR/SQL space) Continue reading “High Performance Analytics”

Amazon goes HPC and GPU: Dirk E to revise his R HPC book

Looking south above Interstate 80, the Eastsho...
Image via Wikipedia

Amazon just did a cluster Christmas present for us tech geek lizards- before Google could out doogle them with end of the Betas (cough- its on NDA)

Clusters used by Academic Departments now have a great chance to reduce cost without downsizing- but only if the CIO gets the email.

While Professor Goodnight of SAS / North Carolina University is still playing time sharing versus mind sharing games with analytical birdies – his 70 mill server farm set in Feb last is about to get ready

( I heard they got public subsidies for environment- but thats historic for SAS– taking public things private -right Prof as SAS itself began as a publicly funded project. and that was in the 1960s and they didnt even have no lobbyists as well. )

In realted R news, Dirk E has been thinking of a R HPC book without paying attention to Amazon but would now have to include Amazon

(he has been thinking of writing that book for 5 years, but hey he’s got a day job, consulting gigs with revo, photo ops at Google, a blog, packages to maintain without binaries, Dirk E we await thy book with bated holes.

Whos Dirk E – well http://dirk.eddelbuettel.com/ is like the Terminator of R project (in terms of unpronounceable surnames)

Back to the cause du jeure-

 

From http://aws.amazon.com/ec2/hpc-applications/ but minus corporate buzz words.

 

Unique to Cluster Compute and Cluster GPU instances is the ability to group them into clusters of instances for use with HPC

applications. This is particularly valuable for those applications that rely on protocols like Message Passing Interface (MPI) for tightly coupled inter-node communication.

Cluster Compute and Cluster GPU instances function just like other Amazon EC2 instances but also offer the following features for optimal performance with HPC applications:

  • When run as a cluster of instances, they provide low latency, full bisection 10 Gbps bandwidth between instances. Cluster sizes up through and above 128 instances are supported.
  • Cluster Compute and Cluster GPU instances include the specific processor architecture in their definition to allow developers to tune their applications by compiling applications for that specific processor architecture in order to achieve optimal performance.

The Cluster Compute instance family currently contains a single instance type, the Cluster Compute Quadruple Extra Large with the following specifications:

23 GB of memory
33.5 EC2 Compute Units (2 x Intel Xeon X5570, quad-core “Nehalem” architecture)
1690 GB of instance storage
64-bit platform
I/O Performance: Very High (10 Gigabit Ethernet)
API name: cc1.4xlarge

The Cluster GPU instance family currently contains a single instance type, the Cluster GPU Quadruple Extra Large with the following specifications:

22 GB of memory
33.5 EC2 Compute Units (2 x Intel Xeon X5570, quad-core “Nehalem” architecture)
2 x NVIDIA Tesla “Fermi” M2050 GPUs
1690 GB of instance storage
64-bit platform
I/O Performance: Very High (10 Gigabit Ethernet)
API name: cg1.4xlarge

.

Sign Up for Amazon EC2

HP goes GPU, Will software people follow

A graphics processing unit on an Nvidia GeForc...
Image via Wikipedia

One more addition to the GPU stack that adds up power when combined with CPU and GPUs. For numeric computing, it may be essential to have GPU- CPU mixed software as almost all hardware people now have offered GPU-CPU products. Maybe software companies can get inspired for new kind of GPU-CPU blade server software again.

Source-

http://www.hpcwire.com/features/HP-Adds-New-HPC-Server-with-GPGPU-Option-104381494.html

But for “true” supercomputing applications, the SL390s G7 is the go-to server. Like its sibling, the SL390s comes with Xeon 5600 processors, but the option to pair the CPUs with up to three on-board NVIDIA “Fermi” 20-series GPUs puts a lot more floating point performance into this design. Customers can choose from either the M2050 or M2070 Tesla GPU modules, the only difference being the amount of graphics memory — 3 GB of GDDR5 for the M2050 versus 6 GB for the M2070. Each GPU module is served by its own PCIe Gen2 x16 channel in order to maximize bandwidth to the graphics chips. At the maximum configuration with all three Fermi GPUs and two Westmere CPUs, a single server delivers on the order of 1 teraflop of double precision performance. “So this is very much a server that has been designed for HPC,” said Turkel.

With GPUs on board, the SL390s fill out a 2U half-width tray, so up to four of these can be packed into a 4U SL6500 chassis. A CPU-only version is also available and takes up just half the space (half-width 1U), enabling twice as many Xeons to occupy the same chassis. This configuration will likely be the server of choice for the majority of HPC setups, given that GPGPU deployment is really just getting started. Pricing on the CPU-only model starts at $2,259.

And

, the ProLiant SL390s G7, provides more raw FLOPS per square inch than any server HP has delivered to date, and is the basis for the 2.4 petaflop TSUBAME 2.0 supercomputer currently being deployed at the Tokyo Institute of Technology.

R on Windows HPC Server

From HPC Wire, the newsletter/site for all HPC news-

Source- Link

PALO ALTO, Calif., Sept. 20 — Revolution Analytics, the leading commercial provider of software and support for the popular open source R statistics language, today announced it will deliver Revolution R Enterprise for Microsoft Windows HPC Server 2008 R2, released today, enabling users to analyze very large data sets in high-performance computing environments.

R is a powerful open source statistics language and the modern system for predictive analytics. Revolution Analytics recently introduced RevoScaleR, new “Big Data” analysis capabilities, to its R distribution, Revolution R Enterprise. RevoScaleR solves the performance and capacity limitations of the R language by with parallelized algorithms that stream data across multiple cores on a laptop, workstation or server. Users can now process, visualize and model terabyte-class data sets at top speeds — without the need for specialized hardware.

“Revolution Analytics is pleased to support Microsoft’s Technical Computing initiative, whose efforts will benefit scientists, engineers and data analysts,” said David Champagne, CTO at Revolution. “We believe the engineering we have done for Revolution R Enterprise, in particular our work on big-data statistics and multicore computing, along with Microsoft’s HPC platform for technical computing, makes an ideal combination for high-performance large scale statistical computing.”

“Processing and analyzing this ‘big data’ is essential to better prediction and decision making,” said Bill Hamilton, director of technical computing at Microsoft Corp. “Revolution R Enterprise for Windows HPC Server 2008 R2 gives customers an extremely powerful tool that handles analysis of very large data and high workloads.”

To learn more about Revolution R Enterprise and its Big Data capabilities, download thewhite paper. Revolution Analytics also has an on-demand webcast, “High-performance analytics with Revolution R and Windows HPC Server,” available online.

AND from Microsoft’s website

http://www.microsoft.com/hpc/en/us/solutions/hpc-for-life-sciences.aspx

REvolution R Enterprise »

REvolution Computing

REvolution R Enterprise is designed for both novice and experienced R users looking for a production-grade R distribution to perform mission critical predictive analytics tasks right from the desktop and scale across multiprocessor environments. Featuring RPE™ REvolution’s R Productivity Environment for Windows.

Of course R Enterprise is available on Linux but on Red Hat Enterprise Linux- it would be nice to see Amazom Machine Images as well as Ubuntu versions as well.

An Amazon Machine Image (AMI) is a special type of virtual appliance which is used to instantiate (create) a virtual machine within the Amazon Elastic Compute Cloud. It serves as the basic unit of deployment for services delivered using EC2.[1]

Like all virtual appliances, the main component of an AMI is a read-only filesystem image which includes an operating system (e.g., Linux, UNIX, or Windows) and any additional software required to deliver a service or a portion of it.[2]

The AMI filesystem is compressed, encrypted, signed, split into a series of 10MB chunks and uploaded into Amazon S3 for storage. An XML manifest file stores information about the AMI, including name, version, architecture, default kernel id, decryption key and digests for all of the filesystem chunks.

An AMI does not include a kernel image, only a pointer to the default kernel id, which can be chosen from an approved list of safe kernels maintained by Amazon and its partners (e.g., RedHat, Canonical, Microsoft). Users may choose kernels other than the default when booting an AMI.[3]

[edit]Types of images

  • Public: an AMI image that can be used by any one.
  • Paid: a for-pay AMI image that is registered with Amazon DevPay and can be used by any one who subscribes for it. DevPay allows developers to mark-up Amazon’s usage fees and optionally add monthly subscription fees.

SAS/Blades/Servers/ GPU Benchmarks

Just checked out cool new series from NVidia servers.

Now though SAS Inc/ Jim Goodnight thinks HP Blade Servers are the cool thing- the GPU takes hardware high performance computing to another level. It would be interesting to see GPU based cloud computers as well – say for the on Demand SAS (free for academics and students) but which has had some complaints of being slow.

See this for SAS and Blade Servers-

http://www.sas.com/success/ncsu_analytics.html

To give users hands-on experience, the program is underpinned by a virtual computing lab (VCL), a remote access service that allows users to reserve a computer configured with a desired set of applications and operating system and then access that computer over the Internet. The lab is powered by an IBM BladeCenter infrastructure, which includes more than 500 blade servers, distributed between two locations. The assignment of the blade servers can be changed to meet shifts in the balance of demand among the various groups of users. Laura Ladrie, MSA Classroom Coordinator and Technical Support Specialist, says, “The virtual computing lab chose IBM hardware because of its quality, reliability and performance. IBM hardware is also energy efficient and lends itself well to high performance/low overhead computing.

Thats interesting since IBM now competes (as owner of SPSS) and also cooperates with SAS Institute

And

http://www.theaustralian.com.au/australian-it/the-world-according-to-jim-goodnight-blade-switch-slashes-job-times/story-e6frgakx-1225888236107

You’re effectively turbo-charging through deployment of many processors within the blade servers?

Yes. We’ve got machines with 192 blades on them. One of them has 202 or 203 blades. We’re using Hewlett-Packard blades with 12 CP cores on each, so it’s a total 2300 CPU cores doing the computation.

Our idea was to give every one of those cores a little piece of work to do, and we came up with a solution. It involved a very small change to the algorithm we were using, and it’s just incredible how fast we can do things now.

I don’t think of it as a grid, I think of it as essentially one computer. Most people will take a blade and make a grid out of it, where everything’s a separate computer running separate jobs.

We just look at it as one big machine that has memory and processors all over the place, so it’s a totally different concept.

GPU servers can be faster than CPU servers, though , Professor G.




Source-

http://www.nvidia.com/object/preconfigured_clusters.html

TESLA GPU COMPUTING SOLUTIONS FOR DATA CENTERS
Supercharge your cluster with the Tesla family of GPU computing solutions. Deploy 1U systems from NVIDIA or hybrid CPU-GPU servers from OEMs that integrate NVIDIA® Tesla™ GPU computing processors.

When compared to the latest quad-core CPU, Tesla 20-series GPU computing processors deliver equivalent performance at 1/20th the power consumption and 1/10th the cost. Each Tesla GPU features hundreds of parallel CUDA cores and is based on the revolutionary NVIDIA® CUDA™ parallel computing architecture with a rich set of developer tools (compilers, profilers, debuggers) for popular programming languages APIs like C, C++, Fortran, and driver APIs like OpenCL and DirectCompute.

NVIDIA’s partners provide turnkey easy-to-deploy Preconfigured Tesla GPU clusters that are customizable to your needs. For 3D cloud computing applications, our partners offer the Tesla RS clusters that are optimized for running RealityServer with iray.

Available Tesla Products for Data Centers:
– Tesla S2050
– Tesla M2050/M2070
– Tesla S1070
– Tesla M1060

Also I liked the hybrid GPU and CPU

And from a paper on comparing GPU and CPU using Benchmark tests on BLAS from a Debian- Dirk E’s excellent blog

http://dirk.eddelbuettel.com/blog/

Usage of accelerated BLAS libraries seems to shrouded in some mystery, judging from somewhat regularly recurring requests for help on lists such as r-sig-hpc(gmane version), the R list dedicated to High-Performance Computing. Yet it doesn’t have to be; installation can be really simple (on appropriate systems).

Another issue that I felt needed addressing was a comparison between the different alternatives available, quite possibly including GPU computing. So a few weeks ago I sat down and wrote a small package to run, collect, analyse and visualize some benchmarks. That package, called gcbd (more about the name below) is now onCRAN as of this morning. The package both facilitates the data collection for the paper it also contains (in the vignette form common among R packages) and provides code to analyse the data—which is also included as a SQLite database. All this is done in the Debian and Ubuntu context by transparently installing and removing suitable packages providing BLAS implementations: that we can fully automate data collection over several competing implementations via a single script (which is also included). Contributions of benchmark results is encouraged—that is the idea of the package.

And from his paper on the same-

Analysts are often eager to reap the maximum performance from their computing platforms.

A popular suggestion in recent years has been to consider optimised basic linear algebra subprograms (BLAS). Optimised BLAS libraries have been included with some (commercial) analysis platforms for a decade (Moler 2000), and have also been available for (at least some) Linux distributions for an equally long time (Maguire 1999). Setting BLAS up can be daunting: the R language and environment devotes a detailed discussion to the topic in its Installation and Administration manual (R Development Core Team 2010b, appendix A.3.1). Among the available BLAS implementations, several popular choices have emerged. Atlas (an acronym for Automatically Tuned Linear Algebra System) is popular as it has shown very good performance due to its automated and CPU-speci c tuning (Whaley and Dongarra 1999; Whaley and Petitet 2005). It is also licensed in such a way that it permits redistribution leading to fairly wide availability of Atlas.1 We deploy Atlas in both a single-threaded and a multi-threaded con guration. Another popular BLAS implementation is Goto BLAS which is named after its main developer, Kazushige Goto (Goto and Van De Geijn 2008). While `free to use’, its license does not permit redistribution putting the onus of con guration, compilation and installation on the end-user. Lastly, the Intel Math Kernel Library (MKL), a commercial product, also includes an optimised BLAS library. A recent addition to the tool chain of high-performance computing are graphical processing units (GPUs). Originally designed for optimised single-precision arithmetic to accelerate computing as performed by graphics cards, these devices are increasingly used in numerical analysis. Earlier criticism of insucient floating-point precision or severe performance penalties for double-precision calculation are being addressed by the newest models. Dependence on particular vendors remains a concern with NVidia’s CUDA toolkit (NVidia 2010) currently still the preferred development choice whereas the newer OpenCL standard (Khronos Group 2008) may become a more generic alternative that is independent of hardware vendors. Brodtkorb et al. (2010) provide an excellent recent survey. But what has been lacking is a comparison of the e ective performance of these alternatives. This paper works towards answering this question. By analysing performance across ve di erent BLAS implementations|as well as a GPU-based solution|we are able to provide a reasonably broad comparison.

Performance is measured as an end-user would experience it: we record computing times from launching commands in the interactive R environment (R Development Core Team 2010a) to their completion.

And

Basic Linear Algebra Subprograms (BLAS) provide an Application Programming Interface
(API) for linear algebra. For a given task such as, say, a multiplication of two conformant
matrices, an interface is described via a function declaration, in this case sgemm for single
precision and dgemm for double precision. The actual implementation becomes interchangeable
thanks to the API de nition and can be supplied by di erent approaches or algorithms. This
is one of the fundamental code design features we are using here to benchmark the di erence
in performance from di erent implementations.
A second key aspect is the di erence between static and shared linking. In static linking,
object code is taken from the underlying library and copied into the resulting executable.
This has several key implications. First, the executable becomes larger due to the copy of
the binary code. Second, it makes it marginally faster as the library code is present and
no additional look-up and subsequent redirection has to be performed. The actual amount
of this performance penalty is the subject of near-endless debate. We should also note that
this usually amounts to only a small load-time penalty combined with a function pointer
redirection|the actual computation e ort is unchanged as the actual object code is identi-
cal. Third, it makes the program more robust as fewer external dependencies are required.
However, this last point also has a downside: no changes in the underlying library will be
reected in the binary unless a new build is executed. Shared library builds, on the other
hand, result in smaller binaries that may run marginally slower|but which can make use of
di erent libraries without a rebuild.

Basic Linear Algebra Subprograms (BLAS) provide an Application Programming Interface(API) for linear algebra. For a given task such as, say, a multiplication of two conformantmatrices, an interface is described via a function declaration, in this case sgemm for singleprecision and dgemm for double precision. The actual implementation becomes interchangeablethanks to the API de nition and can be supplied by di erent approaches or algorithms. Thisis one of the fundamental code design features we are using here to benchmark the di erencein performance from di erent implementations.A second key aspect is the di erence between static and shared linking. In static linking,object code is taken from the underlying library and copied into the resulting executable.This has several key implications. First, the executable becomes larger due to the copy ofthe binary code. Second, it makes it marginally faster as the library code is present andno additional look-up and subsequent redirection has to be performed. The actual amountof this performance penalty is the subject of near-endless debate. We should also note thatthis usually amounts to only a small load-time penalty combined with a function pointerredirection|the actual computation e ort is unchanged as the actual object code is identi-cal. Third, it makes the program more robust as fewer external dependencies are required.However, this last point also has a downside: no changes in the underlying library will bereected in the binary unless a new build is executed. Shared library builds, on the otherhand, result in smaller binaries that may run marginally slower|but which can make use ofdi erent libraries without a rebuild.

And summing up,

reference BLAS to be dominated in all cases. Single-threaded Atlas BLAS improves on the reference BLAS but loses to multi-threaded BLAS. For multi-threaded BLAS we nd the Goto BLAS dominate the Intel MKL, with a single exception of the QR decomposition on the xeon-based system which may reveal an error. The development version of Atlas, when compiled in multi-threaded mode is competitive with both Goto BLAS and the MKL. GPU computing is found to be compelling only for very large matrix sizes. Our benchmarking framework in the gcbd package can be employed by others through the R packaging system which could lead to a wider set of benchmark results. These results could be helpful for next-generation systems which may need to make heuristic choices about when to compute on the CPU and when to compute on the GPU.

Source – DirkE’paper and blog http://dirk.eddelbuettel.com/papers/gcbd.pdf

Quite appropriately-,

Hardware solutions or atleast need to be a part of Revolution Analytic’s thinking as well. SPSS does not have any choice anymore though 😉

It would be interesting to see how the new SAS Cloud Computing/ Server Farm/ Time Sharing facility is benchmarking CPU and GPU for SAS analytics performance – if being done already it would be nice to see a SUGI paper on the same at http://sascommunity.org.

Multi threading needs to be taken care automatically by statistical software to optimize current local computing (including for New R)

Acceptable benchmarks for testing hardware as well as software need to be reinforced and published across vendors, academics  and companies.

What do you think?