ABSTRACT In distributed Software Defined Networking (SDN), multiple controllers need to maintain a consistent view of the network state among themselves using consensus algorithms, introducing additional communication overhead and network delay, especially in large‐scale networks. Therefore, optimizing controller placement presents significant challenges, as it must account not only for the delay between switches and controllers but also for the delay introduced by consensus algorithms. Additionally, SDN controllers have limited capacity in terms of the number of switches they can manage and the network events they can process. Improper placement of controllers can lead to longer message processing times, increased queuing delays, or even controller failures. Thus, achieving balanced workloads among controllers is essential. This study introduces and validates a practical Flow Setup Time (FST) model to measure controller response times. We proposed an advanced multi‐objective optimization approach that incorporates the Variance of Load Balancing (VOLB), to determine the optimal placements of controllers and datastore nodes involved in processing consensus algorithms. Furthermore, we applied this optimization method to different types of real networks from the Internet Topology Zoo dataset. Based on experimental findings, we identified key factors to consider when selecting optimal placement strategies, including the trade‐offs between the number of controllers, the number of datastore nodes, FST, and VOLB.

The increasing complexity of machine learning (ML) and artificial intelligence (AI) applications necessitates efficient GPU resource management in distributed environments such as Kubernetes. Conventional one-to-one GPU mapping, which allocates a single G PU to a single container, often results in the underutilization of these critical resources. Our study introduces an approach that leverages KubeRay and time slicing to enable dynamic GPU sharing among multiple concurrent workloads, significantly improving memory utilization and overall response times. Our findings reveal that while memory efficiency is notably enhanced, the proposed method incurs longer task completion times due to the overhead associated with managing distributed tasks. Specifically, we observed an average increase in task completion times of approximately 74.43 % with two parallel workloads. For three parallel workloads, the average increase in completion times was approximately 158.4 %. This study reveals the trade-offs between improved resource utilization and execution time, highlighting the need for future research to optimize these mechanisms in Kubernetes-based ML operations.

Various animals, including humans, have been suggested to perform Bayesian inferences to handle noisy, time-varying external information. In performing Bayesian inference by the brain, the prior distribution must be acquired and represented by sampling noisy external inputs. However, the mechanism by which neural activities represent such distributions has not yet been elucidated. Our findings reveal that networks with modular structures, composed of fast and slow modules, are adept at representing this prior distribution, enabling more accurate Bayesian inferences. Specifically, the modular network that consists of a main module connected with input and output layers and a sub-module with slower neural activity connected only with the main module outperformed networks with uniform time scales. Prior information was represented specifically by the slow sub-module, which could integrate observed signals over an appropriate period and represent input means and variances. Accordingly, the neural network could effectively predict the time-varying inputs. Furthermore, by training the time scales of neurons starting from networks with uniform time scales and without modular structure, the above slow-fast modular network structure and the division of roles in which prior knowledge is selectively represented in the slow sub-modules spontaneously emerged. These results explain how the prior distribution for Bayesian inference is represented in the brain, provide insight into the relevance of modular structure with time scale hierarchy to information processing, and elucidate the significance of brain areas with slower time scales.

Infrastructure as Code (IaC) has become the de-facto standard method for managing cloud resources. Just like general source code (e.g., Java, etc.), infrastructure code also has numerous bugs so it needs to be tested. While several testing frameworks for IaC for cloud systems have been developed in practice, researchers have paid little attention to their testing. This study presents an empirical investigation of the use of tests for IaC for cloud systems. Our empirical results show that (i) 55.2% of the repositories using Terratest have at least one server infrastructure test; (ii) developers often maintain server infrastructure tests (1.7%-11.3% commits out of all the commits); (iii) many repositories have tests for system functionality (28%), deployment (20%), and configuration (17%).

The optimal path varies depending on the communication characteristics of each application. However, existing routing protocols such as BGP and OSPF do not take this fact into account. Although Software Defined Networking (SDN) has been considered as a possible solution to this problem, SDN technologies relying on a centralized controller has scalability issues. SRv6, which is a source routing protocol that enables SDN, handles routing decisions in a decentralized manner and is expected to scale better than previous SDN technologies.This paper proposes Acar, an adaptive routing system using SRv6 that adaptively controls routing by considering the bandwidth requirements of applications and the link utilization of the network. We conducted experiments on a virtual network and demonstrated that Acar achieves better load balancing between links and higher throughput compared to ECMP.

Federated learning trains a model on a centralized server using datasets distributed over a massive amount of edge devices. Since federated learning does not send local data from edge devices to the server, it preserves data privacy. It transfers the local models from edge devices instead of the local data. However, communication costs are frequently a problem in federated learning. This paper proposes a novel method to reduce the required communication cost for federated learning by transferring only top updated parameters in neural network models. The proposed method allows adjusting the criteria of updated parameters to trade-off the reduction of communication costs and the loss of model accuracy. We evaluated the proposed method using diverse models and datasets and found that it can achieve comparable performance to transfer original models for federated learning. As a result, the proposed method has achieved a reduction of the required communication costs around 90% when compared to the conventional method for VGG16. Furthermore, we found out that the proposed method is able to reduce the communication cost of a large model more than of a small model due to the different threshold of updated parameters in each model architecture.

Animals make efficient probabilistic inferences based on uncertain and noisy information from the outside environment. It is known that probabilistic population codes, which have been proposed as a neural basis for encoding probability distributions, allow general neural networks (NNs) to perform near-optimal point estimation. However, the mechanism of sampling-based probabilistic inference has not been clarified. In this study, we trained two types of artificial NNs, feedforward NN (FFNN) and recurrent NN (RNN), to perform sampling-based probabilistic inference. Then we analyzed and compared their mechanisms of sampling. We found that sampling in RNN was performed by a mechanism that efficiently uses the properties of dynamical systems, unlike FFNN. In addition, we found that sampling in RNNs acted as an inductive bias, enabling a more accurate estimation than in maximum a posteriori estimation. These results provide important arguments for discussing the relationship between dynamical systems and information processing in NNs.

This paper presents a summary of the Masked Face Recognition Competitions (MFR) held within the 2021 International Joint Conference on Biometrics (IJCB 2021). The competition attracted a total of 10 participating teams with valid submissions. The affiliations of these teams are diverse and associated with academia and industry in nine different countries. These teams successfully submitted 18 valid solutions. The competition is designed to motivate solutions aiming at enhancing the face recognition accuracy of masked faces. Moreover, the competition considered the deployability of the proposed solutions by taking the compactness of the face recognition models into account. A private dataset representing a collaborative, multisession, real masked, capture scenario is used to evaluate the submitted solutions. In comparison to one of the topperforming academic face recognition solutions, 10 out of the 18 submitted solutions did score higher masked face verification accuracy.

Federated learning trains a model on a centralized server using datasets distributed over a large number of edge devices. Applying federated learning ensures data privacy because it does not transfer local data from edge devices to the server. Existing federated learning algorithms assume that all deployed models share the same structure. However, it is often infeasible to distribute the same model to every edge device because of hardware limitations such as computing performance and storage space. This paper proposes a novel federated learning algorithm to aggregate information from multiple heterogeneous models. The proposed method uses weighted average ensemble to combine the outputs from each model. The weight for the ensemble is optimized using black box optimization methods. We evaluated the proposed method using diverse models and datasets and found that it can achieve comparable performance to conventional training using centralized datasets. Furthermore, we compared six different optimization methods to tune the weights for the weighted average ensemble and found that tree parzen estimator achieves the highest accuracy among the alternatives.

Empirical Dynamic Modeling (EDM) is a nonlinear time series causal inference framework. The latest implementation of EDM, cppEDM, has only been used for small datasets due to computational cost. With the growth of data collection capabilities, there is a great need to identify causal relationships in large datasets. We present mpEDM, a parallel distributed implementation of EDM optimized for modern GPU-centric supercomputers. We improve the original algorithm to reduce redundant computation and optimize the implementation to fully utilize hardware resources such as GPUs and SIMD units. As a use case, we run mpEDM on AI Bridging Cloud Infrastructure (ABCI) using datasets of an entire animal brain sampled at single neuron resolution to identify dynamical causation patterns across the brain. mpEDM is 1,530× faster than cppEDM and a dataset containing 101,729 neuron was analyzed in 199 seconds on 512 nodes. This is the largest EDM causal inference achieved to date.

Running neural network models on edge devices is attracting much attention by neural network researchers since edge computing technology is becoming more powerful than ever. However, deploying large neural network models on edge devices is challenging due to the limitation in available computing resources and storage space. Therefore, model compression techniques have been recently studied to reduce the model size and fit models on resource-limited edge devices. Compressing neural network models reduces the size of a model, but also degrades the accuracy of the model since it reduces the precision of weights in the model. Consequently, a retraining method is required to recover the accuracy of compressed models. Most existing retraining methods require the original labeled training datasets to retrain the models, but labeling is a time-consuming process. In particular, we cannot always access the original labeled datasets because of privacy policies and license limitations. In this paper, we propose a method to retrain a compressed neural network model with an unlabeled dataset that is different from the original labeled dataset. We compress the neural network model using quantization to decrease the size of the model. Subsequently, the compressed model is retrained by our proposed retraining method without using a labeled dataset to recover the accuracy of the model. We compared the proposed retraining method against the conventional retraining. The proposed method reduced the size of VGG-16 and ResNet-50 by 81.10% and 52.45%, respectively without significant accuracy loss. In addition, our proposed retraining method is clearly faster than the conventional retraining method.

Data centers are centralized facilities where computing and networking hardware are aggregated to handle large amounts of data and computation. In a data center, computing resources such as CPU and memory are usually managed by a resource manager. The resource manager accepts resource requests from users and allocates resources to their applications. A commonly known problem in resource management is that users often request more resources than their applications actually use. This leads to the degradation of overall resource utilization in a data center. This paper aims to improve resource utilization in data centers by predicting the required resource for each application. We designed and implemented a neural network model based on Long Short-Term Memory (LSTM) to predict more efficient resource allocation for a job based on historical data. Our model has two LSTM layers each of which learns the relationship between: (1) allocation and usage, and (2) CPU and memory. We used Googles cluster-usage trace, which contains a trace of resource allocation and usage for each job executed on a Google data center, to train our neural network. Googles cluster scheduler simulator was used to evaluate our proposed method. Our simulation indicated that the proposed method improved the CPU utilization and memory utilization by 10.71% and 47.36%, respectively, compared to a conventional resource manager. Moreover, we discovered that increasing the memory cell size of our LSTM model improves the accuracy of the prediction in return for longer training time.

Multi-tenant data centers for cloud computing require the deployment of virtual private networks for tenants in an on-demand manner, providing isolation and security between tenants. To address these requirements, network virtualization techniques such as encapsulation and tunneling have been widely used. However, these approaches inherently incur processing overhead on end-points (such as the host hypervisor), reducing the effective throughput for the tenant virtual network compared to the native network. This problem is exacerbated with increases in line rates, now exceeding 10Gbps. In this paper, we introduce PARES (PAcket REwriting on SDN), a novel technique which uses the packet rewriting feature of SDN switches to provide multi-tenancy in data center networks at edge switches, thereby reducing the load on end-point hypervisors and improving the throughput, compared to tunneling. Experiments in an SDN testbed show that our proposed data center arhictecture with PARES achieves near line-rate multi-tenancy virtualization with 10Gbps links (compared to 20% of line-rate for VXLAN tunneling), without incurring processing overhead at end-point hypervisors or guest servers. Additionally, the paper evaluates the scalability of PARES for ARP protocol handling and with respect to number of SDN flow entries.

Since the Internet is an aggregation of multiple ASes (Autonomous Systems), congestion control and utilization are not globally optimized. For example, it is not uncommon that a direct shortest route with low latency delivers less bandwidth than an alternative, long and roundabout route. Previous research has shown that geospatially distributed computing instances in commercial clouds offer users an opportunity to deploy relay points to detour potentially congested ASes, and as a means to diversify paths to increase overall bandwidth and reliability. Such opportunity comes with a cost, as cloud-routed paths incur cost of not only provisioning of computing resources, but also for additional traffic to/from Internet. Well-established protocols, such as TCP, were created based on assumption of single end-point to end-point transfer; nonetheless, current computing devices have multiple end-points, and the increasing availability of overlay networks allows multiplexing multiple virtual network flows into a single physical network interface. In this paper, we empirically evaluate the extent to which using cloud paths to transfer data in parallel with the default Internet path can improve the end-to-end bandwidth in bulk data transfers. In our evaluation, we consider single-stream and multi-stream TCP transfers across one or more paths. Moreover, we suggest an application level design pattern that takes advantage of this improved aggregate bandwidth to reduce data transfer times.

Similar to Virtualization, Linux Containers (LXC) provides high-performance, lightweight computing resource allocation and isolation. Each LXC container has a resource overhead smaller than that of a virtual machine, leading to significantly lower container migration time and making frequent container placement modification a viable optimization technique. Traditional container scheduling mechanisms do not leverage this property of LXC. Generally, a scheduler tries to find the most optimal placement for a new container, the allocated host then executes the scheduled container until the end of the container's life cycle. This strategy works fine for short-lived containers. With a long-lived container such as a server process becoming more and more common, and the container placement calculated at the beginning of the execution may not remain optimal during the container's lifetime, since the other containers are moving in and out of the cluster. This research proposes container rebalancing, a novel scheduling mechanism with a rebalancing process working alongside a scheduling process. The container rebalancing method increases LXC cluster utilization while maintaining minimal interference with the scheduling process. This is done by continuously modifying container placement, by using the rebalancing process, in order to load-balance utilization of each host in the LXC cluster. LXC cluster simulation driven by Google's cluster data is used to verify the feasibility of container rebalancing. Simulation results show an observable increase in container scheduled rate and cluster utilization with no drawback, suggesting that container rebalancing is a promising method.

Summary The Pacific Rim Application and Grid Middleware Assembly (PRAGMA) is an international community of researchers that actively collaborate to address problems and challenges of common interest in eScience. The PRAGMA Experimental Network Testbed (PRAGMA‐ENT) was established with the goal of constructing an international software‐defined network (SDN) testbed to offer the necessary networking support to the PRAGMA cyberinfrastructure. PRAGMA‐ENT is isolated, and PRAGMA researchers have complete freedom to access network resources to develop, experiment, and evaluate new ideas without the concerns of interfering with production networks. In the first phase, PRAGMA‐ENT focused on establishing an international L2 backbone. With support from the Florida Lambda Rail, Internet2, PacificWave, Japan Gigabit Network, and TaiWan Advanced Research and Education Network, PRAGMA‐ENT backbone connects openflow‐enabled switches at University of Florida, University of California, San Diego, Nara Institute of Science and Technology (Japan), Osaka University (Japan), National Institute of Advanced Industrial Science and Technology (Japan), and National Applied Research Laboratories (Taiwan). The second phase of PRAGMA‐ENT consisted of an evaluation of technologies for the control plane that enables multiple experiments (ie, OpenFlow controllers) to coexist. Preliminary experiments with FlowVisor revealed some limitations leading to the development of a new approach, called AutoVFlow. This paper describes our experience in the establishment of PRAGMA‐ENT backbone (with international L2 links), its current status, and plans for the control plane. Discussion of preliminary application ideas, including optimization of routing control; multipath routing control; extending the backbone using overlay network; and remote visualization are also discussed.

Summary Multipath transmission control protocol(TCP), or MPTCP, is a widely‐researched mechanism that allows a single application‐level connection to be split to more than 1 TCP stream and, consequently, more than 1 network interface, as opposed to the traditional TCP/IP model. Being a transport layer protocol, MPTCP can easily interact between the application using it and the network supporting it. However, MPTCP does not have control of its own route. Default IP routing behavior generally takes all traffic through the shortest or best metric path. However, this behavior may actually cause paths to collide with each other, creating contention for bandwidth in a number of edges. This can result in a bottleneck that limits the throughput of the network. Therefore, a multipath routing mechanism is necessary to ensure smooth operation of MPTCP. We created smoc, a simple multipath OpenFlow controller, that uses only topology information of the network to avoid collision where possible. Evaluation of smoc in a virtual local‐area and a physical wide–area software‐defined networks showed favorable results as smoc provided better performance than simple or spanning tree–routing mechanisms.

We demonstrate high-speed data transfer GridFTP using a multipath control mechanism with SDN (Software-Defined Networking). GridFTP is a typical tool that has been developed and widely used for bulk data transfer over a wide area network in the field. GridFTP supports a parallel high-speed data transfer scheme using multiple TCP streams. However, one of the shortest paths is used solely for data transfer in the default IP routing while there are multiple network paths (multipath) exist between widely-distributed sites. In this study, we propose a system that distributes the parallel TCP streams of GridFTP into multiple network paths by a traffic engineering technique brought by SDN. Our proposed system has achieved approximately 20% better performance than the conventional method in the best case in a global-scale real enviroment.

A large amount of scientific data needs to be transferred from one site to another as fast as possible in the computational science fields. High-speed data transfer between sites is very important, especially in the Grid computing field, GridFTP has been widely used for bulk data transfer over a wide area network. GridFTP achieves greater performance by supporting parallel TCP streams. Using parallel TCP streams improves the throughput of slow-start algorithms and lossy networks even on a single path. This research proposes a traffic engineering technique that increases the data transfer performance by using multiple paths simultaneously for the parallel TCP streams. For this purpose, we use Software-Defined Network (SDN) technology and its implementation, OpenFlow. This paper presents the design and implementation of the proposed system. Our performance evaluation demonstrates that our proposed system can accelerate GridFTP Transfer in both virtual and real global-scale environments.

With the constant increase in the number and variety of small molecule chemical compounds, drug discovery is becoming a very resource intensive endeavor. Performing molecular simulations of ligand-protein binding by virtual screening has become an integral part of the discovery process. Cloud computing is an efficient choice to execute these large-scale screenings, given that large compute allocations are not accessible to many researchers. This research focused on developing a multi-site cloud environment that combines small allocations of virtual machines in multiple locations connected through a virtual networking system (ViNe), and compared two parallelization approaches: Message Passing Interface (MPI) and MapReduce using Hadoop. Virtual screenings were conducted using DOCK, a protein-ligand molecular interaction simulation program. Multiple DOCK test simulations through MPI and Hadoop were run to assess the performance and flexibility of the environment. These tests indicated that MPI and MapReduce offer comparable scalability performance, and that network latency has a significant influence on low accuracy simulations. Furthermore, differences in performance at individual cloud resource sites were reduced on average because of the larger combined pool of resources. This project prototyped and assessed a fully functional multi-site cloud environment for virtual screenings, which can be used to guide small laboratories in deploying their own cloud-based screenings.

Bandwidth and latency are two major factors that contribute the most to network application performance. Between each pair of switches in a network, there may be multiple paths connecting them. Each path has different properties because of multiple factors. Traditional shortest-path routing does not take this knowledge into consideration and may result in sub-optimal performance of applications and underutilization of network. We proposed a concept of "bandwidth and latency aware routing". The idea is that we could improve overall performance of the network by separating application into bandwidth-oriented and latency-oriented application and allocate different route for each type of application accordingly. We also proposed a design of this network system implemented using Open Flow. Routes are calculated from monitored information using Dijkstra algorithm and its variation. To support our design, we show a use case in which our design performs better than traditional routing as well as evaluation results.

In the era of cloud computing, data centers that accommodate a series of user-requested jobs with a diversity of resource usage pattern need to have the capability of efficiently distributing resources to each user job, based on individual resource usage patterns. In particular, for high-performance computing as a cloud service which allows many users to benefit from a large-scale computing system, a new framework for resource management that treats not only the CPU resources, but also the network resources in the data center is essential. In this paper, an SDN-enhanced JMS that efficiently handles both network and CPU resources and as a result accelerates the execution time of user jobs is introduced as a building block technology for such a HPC cloud. Our evaluation shows that the SDN-enhanced JMS efficiently leverages the fat-tree interconnect of cluster systems running behind the cloud to suppress the collision of communications generated by different jobs.

In recent years, much attention has been paid to wide area distributed storages to backup data remotely and ensure that business processes can continue in terms of disaster recovery. In the 'distcloud' project, authors have been involved in the research of wide area distributed storage by clustering many computer resources located in geographically distributed areas, where the number of sites is more than 2 (N>2). The storage supports a shared single POSIX file system so that LDLM (Long Distance Live Migration) of VMs (Virtual Machines) works well between multiple sites. We introduce the concept and basic architecture of the wide area distributed storage and its technical improvement for LDLM. We describe the result of our experiment, that is, 1) Nation Wide Live Migration (about 500Km) in Japan, and 2) Transpacific Live Migration (over 24,000Km). We show the technical benefit of the current implementation and discuss suitable applications and remaining issues for further research topics.

Recently, the concept of Software-Defined Network (SDN), which allows us to administer and configure a network in a centralized and software-programming manner, has gathered network engineers' and researchers' attention rapidly. In particular, the expectation and concern to OpenFlow as an implementation of the SDN is remarkable. As a result, research activities, which include prototyping, implementation, demonstration and experiments, conducted over OpenFlow networks have been a worldwide tendency. In such research activities, however, the difficulty in understanding network topology, traffic amount and an actual path of a network flow on the OpenFlow network, and the intricacies in debugging software designed for OpenFlow are serious problems in the development process of OpenFlow controller. This research aims to realize a visualization software that facilitates researchers to perform OpenFlow controller development and demonstration experiments performed on an actual OpenFlow network. In this paper, the authors summarize the achievement of their research work in progress as well as the future direction.

Advances in sequencing technology have resulted in an exponential increase in the availability of protein sequence information. In order to fully utilize information, it is important to translate the primary sequences into high-resolution tertiary protein structures. MODELLER is a leading homology modeling method that produces high quality protein structures. In this study, the function of MODELLER was expanded by configuring and deploying it on a parallel grid computing platform using a custom four-step workflow. The workflow consisted of template selection through a protein BLAST algorithm, target-template protein sequence alignment, distribution of model generation jobs among the compute clusters, and final protein model optimization. To test the validity of this workflow, we used the Dual Specificity Phosphatase (DSP) protein family, which shares high homology among each other. Comparison of the DSP member SSH-2 with its model counterpart revealed a minimal 1.3% difference in output energy scores. Furthermore, the Dali Pair wise Comparison Program demonstrated a 98% match among amino acid features and a Z-score of 26.6 indicating very significant similarities between the model and actual protein structure. After confirming the accuracy of our workflow, we generated 23 previously unknown DSP family protein structure models. Over 40,000 models were generated 30 times faster than conventional computing. Virtual receptor-ligand screening results of modeled protein DSP21 were compared with two known structures that had either higher or lower structural homology to DSP21. There was a significant difference (p!0.001) between the average ligand ranking discrepancy of a more homologous protein pair and a less homologous protein pair, suggesting that the protein models generated were sufficiently accurate for virtual screening. These results demonstrate the accuracy and usability of a grid-enabled MODELLER program and the increased efficiency of processing protein structure models. This workflow will help increase the speed of future drug development pipelines.

In 2005, Osaka University, in Japan, started an international educational program called, Pacific Rim International University (PRIUS), on top of the Pacific Rim Application and Grid Middleware Assembly (PRAGMA) research framework. The PRIUS framework is based on and similar to that of the PRIME program at the University of California San Diego. Through the PRIUS program, Osaka University has explored a new structure of higher education for graduate students by combining lectures given by PRAGMA researchers and scientists as well as internship abroad opportunities to PRAGMA member institutions and universities. In this paper, we describe the goals and framework of the PRIUS program and discuss issues for the improvement of PRIUS. We also present two examples of interns' achievements as well as other educational effects brought through collaboration with PRAGMA.

SHP-2 is a protein tyrosine phosphatase (PTP) that plays an important role in many cellular functions such as development, growth, and death; thus SHP-2 has been hypothesized to play an important role in various diseases such as diabetes, neurodegeneration, and cancer. The importance of the individual roles of different PTPs is not well understood and this is complicated by the lack of specific inhibitors. In this study, we have utilized the multi-institutional PRAGMA Grid computation resources to virtually screen the ZINC 7 database using virtual docking software DOCK 6.2. Preliminary results suggest several SHP-2 specific inhibitors that can be further tested and validated under laboratory conditions. Complications during these multiple, virtual screenings on the grid as well as potential improvements are also discussed. These findings have future clinical significance in the creation of new drug therapies for the treatment of different diseases.

The slingshot-2 (SSH-2) protein plays a significant role in different cell functions such as growth and movement. SSH-2 is a phosphatase protein that belongs to a unique class of enzymes called dual specificity phosphatases (DSP) that target the phosphothreonine and phosphotyrosine residues of mitogen-activated protein (MAP) kinases, which regulate cell growth. Because of this, it is of great interest to find specific inhibitors of DSPs such as SSH-2. Implementing an in silico platform to screen a sizable pool of chemical compounds against SSH-2 on the grid environment with the molecular docking software DOCK 6, several chemical compounds have been identified as potential inhibitors of SSH-2 activity. The issues of performing routine virtual screenings on the grid and possible improvements are also presented. The most promising inhibitor determined from standard and AMBER DOCK screenings was 2-amino-3-phosphonooxy-propanoic acid, which will be verified with wet bench testing.