One of the hindrances to scalable, cross-machine distribution of numerical methods is the communication of shared data, otherwise known as ghost regions. Due to the fact that traditional cross-machine parallelism using messaging passing is a serial process, Amdahl’s Law and communication latency dictate that this type of distributed memory approach will scale poorly. However, if a problem is divided into spatial sub-domains for multicore distribution, ghost regions are no longer necessary because adjacent data is readily available in shared memory where communication times are orders of magnitude shorter. However, for this new paradigm of multicore parallel computing to succeed, shared memory challenges such as thread safety and race conditions must be addressed.

The spatial decomposition of a numerical method such as the lattice Boltzmann method (LBM) creates a number of computational tasks. Multicore distribution of these tasks requires the use of a coordination tool to manage them onto processing cores in a load balanced way. Whilst this could be undertaken using the scatter-gather approach of OpenMP we have instead developed an events-based programming model (H-Dispatch) that displays scalability, memory efficiency and intrinsic load balancing. One of the key aspects of H-Dispatch is the way in which tasks are distributed to threads. Rather than a scatter or push of tasks from the manager to threads, instead threads request tasks as they become free. It is this pull mechanism that enables the use of a single thread per core as threads only request a value when free, thus, there is never more than one task at a time associated with a given enduring thread (and its associated local variable memory).

The LBM has been established in the last 20 years as a powerful numerical method for the simulation of fluid flows, with application in areas such as multiphase, porous media, turbulence and particle suspensions. As an example of the potential of multicore parallel computing, an LBM code was ported to the H-Dispatch programming model for simulation of fluid flow in a porous rock specimen. Figure 1a shows schematically how the LBM domain was decomposed into cubic nodal bundles, as it is these bundles which form the computational tasks which are distributed to cores. The results of the rock permeability test are graphed in Figure 1b as speed-up against the number of cores when a domain of 2003 domain is employed.

By varying the side length of the nodal bundles between 10 and 50 the effect on cache blocking, which is integral to the exploitation of the memory hierarchy in modern computers, and subsequently performance was highlighted. Optimal blocking and scaling were achieved at a side length of 20, where the speed-up and efficiency were approximately 22 and 92%, respectively, on 24 cores. As bundles were increased beyond this size, performance degraded due to the tasks becoming too large for cache which resulted in ‘slow’ communication with RAM for data. The optimum side length of 20 was then transferred to identical problems using a 3003 and 4003 domain (results included in Figure 1b). Near-linear scalability was again achieved and at 24 cores the speed ups were approximately 23 and 22, and the efficiencies were approximately 95% and 92% for 3003 and 4003, respectively. This is an important result, as it suggests that the optimum bundle size for LBM problems can be determined in an a priori fashion for a specific architecture. Further, this research forms an excellent basis for the extension of multicore parallelism to even larger problems on larger shared memory systems, either hardware-based or virtualized.

Geospatial platforms have been used as a coordination layer for some of the most devastating events in recent history: the Earthquake in Haiti (January 2010), the Deepwater Horizon oil spill in the Gulf of Mexico (April 2010) and the Earthquake, Tsunami and Nuclear Crisis in Japan (March 2011). Responses to such events require diverse information and knowledge of location. However, geospatial tools do much more than place locations on a map. They serve as the basis for understanding complex issues and relationships among multiple sources of information visually portrayed through maps.

The map is one of the oldest forms of nonverbal communication; the oldest known maps are preserved on Babylonian clay tablets from about 2300 BCE. In present society, Web Mapping has been widely popularized through mapping APIs and personal global positioning systems (GPS).

Mashups amplify the power of these visualization techniques. The light weight protocols enable to overlay of multiple data sets on top of a web geospatial mapping platforms. Mashups can aid in the development of web applications by giving them better interactivity with maps and associated data, and can aid in the manipulation of blogs, lists, photo and video sharing.
The Intelligent Engineering Systems Laboratory is currently exploring a number of geospatial platforms.

What is Cloud Computing? The answer can often be cloudy (pardon the pun).

The students of my class did a great job answering the question in video form. They are Rory, Pierre, and Rachelle – see below.

Dr. Abel Sanchez Lecturer Executive Director, MIT Geospatial Data Center Research Scientist, AutoID Laboratories Research Lead, Intelligent Engineering Systems Laboratory (IESL) Fellow Ford Foundation, National Research Council of the National Academies Dr. Sanchez holds a Ph.D. from the Massachusetts Institute of Technology (MIT). His areas of specialty include Radio Frequency Identification (RFID), Information Engineering, and Engineering Complex Systems. He teaches graduate courses in Software Construction, Software Engineering, and Software Architecture. For the past eight years, his research has focused on architecting of large scale distributed simulation systems.

Pierre Fuller Ph.D Candidate Research Topic: “3D Spatial Orientation” Research Interests Application of sensors in construction and building operations, Building and construction automation, Real-time building models, Automated creation of building models, Real-time navigation and evacuation strategies for building occupants. Education B.S. in Architecture (Magna Cum Laude) from LTU, 2007 B.S. in Civil & Environmental Engineering (Magna Cum Laude) from LTU, 2007 Masters of Science in Civil & Environmental Engineering from MIT, 2009 Ph.D in Civil & Environmental Engineering at MIT, current

Rory Clune Ph.D Candidate Research Topic: “Structural Design and Analysis Software” Research interests: Computational Design, Structural Optimization, Structural Software Development and Integration Education: B.Eng. (Civil) – University College Cork, Ireland, 2008 S.M., Civil and Environmental Engineering – MIT, 2010 Ph.D., Civil and Environmental Engineering – MIT, current

The deployment of Natural Language Processing algorithms on a scalable coherent shared-memory architecture enabled the identification of tweets about either spontaneous or scheduled events in real-time. As an example, we experimented with the specific case of music concerts in New York. The utilization of a feature vector constructed as a trigram model that merged tweets with NYC.com information yielded a 99% of accuracy on the detection of musical events using the Support Vector Machine algorithm.

The Intelligent Engineering Systems Laboratory Congratulates Dr. Leonardi, Dr. Holmes, Dr. Williams, and Dr. Tilke

The paper

A Multi-Core Numerical Framework for Characterizing Flow in Oil Reservoirs,
Christopher Leonardi (MIT), David Holmes (James Cook University), John Williams (MIT) and Peter Tilke (Schlumberger-Doll Research Center)

Abstract

This paper presents a numerical framework that enables scalable, parallel execution of engineering simulations on multi-core, shared memory architectures. Distribution of the simulations is done by selective hash-tabling of the model domain which spatially decomposes it into a number of orthogonal computational tasks. These tasks, the size of which is critical to optimal cache blocking and consequently performance, are then distributed for execution to multiple threads using the previously presented task management algorithm, H-Dispatch. Two numerical methods, smoothed particle hydrodynamics (SPH) and the lattice Boltzmann method (LBM), are discussed in the present work, although the framework is general enough to be used with any explicit time integration scheme. The implementation of both SPH and the LBM within the parallel framework is outlined, and the performance of each is presented in terms of speed-up and efficiency. On the 24-core server used in this research, near linear scalability was achieved for both numerical methods with utilization efficiencies up to 95%. To close, the framework is employed to simulate fluid flow in a porous rock specimen, which is of broad geophysical significance, particularly in enhanced oil recovery.

Authors

Christopher R. Leonardi, chrisleo@mit.edu

David W. Holmes, david.holmes1@jcu.edu.au

John R. Williams, jrw@mit.edu

Peter G. Tilke, tilke@slb.com

Of the billions of Radio Frequency Identification (RFID) tags deployed, most tags are used in supply chains where the Electronic Product Code (EPC) and associated business event data are transmitted through RFID global networks. Moreover, it is estimated that Wal-Mart’s in-store implementation will generate about 7 terabytes of RFID data per day [1]. If inadequate security is in place, unauthorized parties could possibly read RFID tags or analyze the EPC traffic. According to a study conducted by the Ponemon Institute, data breaches cost businesses an average of $197 per customer record. Therefore, security and privacy issues are critically important in RFID networks. Companies need to share these data with restricted business partners and, under some conditions, such as product recall, more widely with regulators and non business partners.

However, no security or privacy framework has been chosen as a global RFID networks standard at present. The lack of agreement on a specific security framework in RFID networks is due to the following reasons. First, some RFID networks, such as the supply chain, need to be worldwide and be relatively easy to access. For example, in a supply chain, a RFID tag is read by manufacturing plants, storage warehouses, carrier cross docks, major distribution centers, ports, intermodal terminals, suppliers, transport carrier, a third-party logistics provider and a retail store. It is hard to make mandatory adoption of a particular identity system or a certificate authority for all these widely distributed vendors. Secondly, there are trust issues in RFID networks. For example, a manufacturer sells product to a distributor, and the distributor sells product to a retailer. When the manufacturer hopes to trace his products and
wants to query for particular RFID number’s information from retailers, the retailers need to decide whether they should give access permission to the manufacturer even though they do not have a direct business relationship with the manufacturer. Finally, the business rules for protecting access to RFID data can be complex. The power of an RFID number lies in its ability to give us access to data about the product and its business context. For example, the product could be an F15 Engine GE504 destined for Pakistan, and the associated business information could be that the engine was shipped from Boeing and was part of a Purchase Order (PO) that is stored on http://boeing.SAPsecretsite/secretPlace.aspx. Thus, the resource being accessed, and for which authorization is sought, is potentially much more than an RFID number and the authentication and authorization decisions depend on all of the following data: RFID number, time, Business Event, PO number. What is more, the user might have partial authorization to see some data but not all of it. So any authorization system must look up a business rule that spans a space defined by several parameters, such as EPC, Shipper, Date, PO Context, and Requester Identity. At present there are no interchangeable business rule languages which can express complex business rules such as this.

In order to address these problems, we propose the Interoperable Internet Scale Security Framework, IISS, for RFID Networks, which has the following features:

•    Centralized identity providers are avoided in the security framework due to the wide variety and huge number of the parties involved in the RFID network.

•    Different authentication schemes and identity techniques are supported, so that different parties with different techniques or platforms are interoperable.

•    Provenance information of the RFID data is generated and passed through the security framework to identify the track of particular RFID data for both traceability and authentication reasons. The provenance information in IISS can be used as a proof for authentication to some extent when two parties do not share any identity authorities.

•    Business rules are interchangeable between agents in different domains.

•    A business rule language is formulized in IISS, so that permission to particular parties can be given according to the inference on particular business rules.

The main contribution of IISS is that it provides a solution to security problems in the RFID networks on which multiple partners can rely. Moreover, in RFID networks, IISS is the first scheme that performs authentication based on an aggregation of business rules, enterprise information, and RFID tag information. Finally, IISS provides a way to express complex business rules in RFID networks.

A ‘smart’ or ‘intelligent’ well is considered one of the most advanced types of nonconventional wells. A typical smart well is equipped with a special completion that has packers or sealing elements which allow partitioning of the wellbore, pressure and temperature sensors and downhole inflow control valves (ICV) installed on the production tubing. The sensors allow continuous monitoring of pressure and temperature while the ICVs provide the flexibility of controlling each branch of a multilateral well independently. A smart well can be either a multilateral well where every lateral is controlled by an ICV or a single bore well where each segment is controlled by an ICV. The advantages of smart wells have been demonstrated in practical applications for both single and multiple reservoir production (non-commingled production). Because of their ability to control production from each lateral or segment through ICV adjustment and manipulation, smart wells can mitigate water production by allocating the optimum production rate and therefore increase the ultimate recovery.

Unlike conventional wells where only surface control is used to determine the optimum production rate, optimization of smart wells requires determining the best combination of ICV settings (ICV configuration) that yields the highest recovery factor and hence profit. In the case of commingled production, laterals or branches are in contact with each other within the immediate vicinity of the reservoir. This adds another dimension to the optimization process as one lateral might affect the production of other laterals (in the case of water breakthrough). Reservoir engineering practices follow two approaches in optimizing oil production from smart wells. These approaches are the reactive approach and the proactive approach. The reactive approach is usually achieved on a trial and error basis, making decisions based on current conditions. A series of production tests is made to determine the best ICV configuration. A portable multiphase flow meter (MPFM) is usually used for a faster decision.  The reactive approach mainly corrects for any deviation from the production target; i.e. fast increase in water cut due to heterogeneity. On the other hand, a proactive approach uses an optimization algorithm such as Genetic Algorithm, Simplex, or Swarm Particles to achieve the best ICV configuration that yields maximum oil recovery over a period of time in particular making estimates of future events. The proactive approach also takes advantage of the availability of real time production data which allows for better decisions.

For more information on reactive optimization approach see:

Yeten, B. (2003). Optimum Deployment of Nonconventional Wells. PhD dissertation, Department of Petroleum Engineering, Stanford University, California.

Yeten, B., Durlofsky, L. J., Aziz, K., (2002). Optimization of Smart Well Control, paper SPE/PS-CIM/CHOA 79031 presented at the SPE International Thermal Operations and Heavy Oil Symposium and International Well Technology Conference in Calgary, Alberta, CANADA, 4-7 November

The aggregation of a traditional cluster, via a software layer such as ScaleMP’s vSMP Foundation, into a single virtual system in which applications can directly address all memory (www.scalemp.com).

One approach to developing nature-inspired, network diagnostic tools may be to employ agent-based software architecture. This would follow the general concepts of agent-based modeling, in which a domain (e.g. network, supply chain, Cartesian space) is populated with a large number of ‘agent’ models that interact and evolve via a set of simplistic rules. With a well defined set of rules, the ensemble behavior of agents can describe complex systems in fields such as biology, sociology and economics.

The central tenet of agent-based modeling is common to a number of mathematical models popular in science and engineering. In particular, cellular automata (CA), smoothed particle hydrodynamics (SPH) and lattice Boltzmann methods (LBM) are constructed using simplistic micro/meso-scale models which yield complex Navier-Stokes hydrodynamics at the macro-scale. The development of software based on these models, via the Geonumerics Laboratory , is an area of expertise of the MIT GDC.

As with engineering modeling using these techniques, simulating the behavior of diagnostic agents over large networks may necessitate intensive computations that can become intractable using serial processing. The solution to this problem is to employ parallel computing, which has historically been undertaken using distributed-memory clusters. However, one of the hindrances to scalable, cross-machine distribution of numerical methods is the communication of data between non-coherent memory at synchronization times, which is plagued by the latency of TCP/IP communication. Additionally, software development in this context requires specialist programming skills in message passing (i.e. MPI).

The performance and implementation challenges of distributed memory parallelism can be overcome by utilizing coherent shared memory hardware. This can be realized in two ways, either via a hard-wired array of multi-core processors  or by aggregating traditional clusters, via a software layer, into a single virtual system. Both approaches have relative strengths and weaknesses, however they both achieve the most important requirement for simplifying the development and optimizing the performance of parallel software, which is to allow the data associated with each distributed computational task to be directly addressable by all other tasks.

For more information see:

D. W. Holmes, J. R. Williams, and P. G. Tilke. Smooth particle hydrodynamics simulations of low Reynolds number flows through porous media. International Journal for Numerical and Analytical Methods in Geomechanics, n/a. doi: 10.1002/nag.898, 2010.

D. W. Holmes, J. R. Williams, and P. G. Tilke. An events based algorithm for distributing concurrent tasks on multi-core architectures. Computer Physics Communications, 181(2):341-354, 2010.

IT infrastructures in global corporations are appropriately compared with nervous systems, in which body parts (interconnected datacenters) exchange signals (pieces of data) in order to coordinate actions (data visualization and manipulation). A priori inoffensive perturbations in the operation of the system or the elements composing the infrastructure can lead to catastrophic consequences. Downtime disables the capability of clients reaching the latest versions of the data and/or propagating their individual contributions to other clients, potentially costing millions of dollars to the organization affected. The imperative need of guaranteeing the proper functioning of the system not only forces to pay particular attention to network outages, hot-objects or application defects, but also slows down the deployment of new capabilities, features and equipment upgrades. Under these circumstances, decision cycles for these improvements can be extremely conservative, and be prolonged for years, involving multiple personalities across departments of the organization. Frequently, the solutions adopted are years behind state-of-the-art technologies or phased out compared to leading research on the data infrastructure field.
In the “Datacenter Arena” project the utilization of a large-scale collaborative data infrastructure simulator is proposed, in order to evaluate the impact of “what if” scenarios on the performance, availability and reliability of the system. The goal is to provide datacenter operators a tool that allows understanding and testing the consequences of the deployment of new network topologies, hardware configurations or software applications before they are effective. The simulator is constructed using a multi-layered approach, providing a granularity down to the individual server component or client action, and was validated against the global infrastructure of a leader automobile company.
Modeling and simulation of IT infrastructures has two main advantages compared to the infrastructure profiling approach used in web-scale companies:
Simplicity and Cost:Profiling datacenter behavior requires running processes that measure vast amounts of highly detailed information. As the infrastructure size increases the profiling overhead can degrade system performance and leaving collected data unexploited, unless additional resources are allocated, which increases the overall profiling cost. The simulation platform proposes a simpler yet powerful non-intrusive tool capable not only of reproducing system behavior on a high level, but also predicting the impact of “what if” scenarios for a lower cost.
Unified Perspective: The high complexity of a global infrastructure makes it impossible for a single datacenter operator to understand all the dynamics of the workload, hardware, software or network; in fact, typically individuals do not have detailed knowledge about the system beyond their area of responsibility. The simulation platform serves as a unique “playground” that enables decision makers being on the same page and reaching consensus on system alterations.
For more information, check our Datacenter Arena research papers.