Practice and Experience in Advanced Research Computing, 2021
CloudBank is a cloud access entity founded to enable the computer science research and education ... more CloudBank is a cloud access entity founded to enable the computer science research and education communities to harness the profound computational potential of public clouds. By delivering a set of managed services designed to alleviate common points of friction associated with cloud adoption, Cloudbank serves as an integrated service provider to the research and education community. These services include front-line help desk support, cloud solution consulting, training, account management, cost monitoring and optimization support, and automated billing. CloudBank has a multi-cloud pay-per-use billing model and aims to serve the spectrum of cloud users from novice to advanced.
because the underloaded hosts ask for chunks of work which involve number of threads which is jus... more because the underloaded hosts ask for chunks of work which involve number of threads which is just too small. In this case extreme load imbalance may occur: some hosts nish their part and stay completely idle, while others still have a lot of work that could otherwise be split. For instance, the system may reach a state in which 20 threads execute on one of the machines. The threads are divided into two groups of 10 threads each, where all the threads in each of the groups are communicating intensively with each other, but no communication exists between the groups. Clearly, one of these groups should be sent to one of the underloaded hosts, which may be prevented by the fact that 10 threads is greater than the overloaded threshold parameter of the load-balancing algorithm. Once again, this requires some smart mechanism which will adaptively identify and resolve this problem. We give here several additional parameters and issues that call for further research (or even simple tuning)...
The availability of 64-bit processors enables operating systems the use of the shared address spa... more The availability of 64-bit processors enables operating systems the use of the shared address space (SAS) paradigm where all processes execute in the same address space. This is in contrast to most existing operating systems that use the private address space (PAS) paradigm where each process views the entire space as dedicated to itself. The PAS paradigm yields high overhead for interprocess communication and sharing, whereas the SAS paradigm can reduce the cost of communication and provide simple abstractions to the application level to build variety of communication schemes. Furthermore, the use of the SAS paradigm results in increased system performance. The SAS paradigm can be classi ed into various models, which di er in the way the address space and protection domain are managed. In this paper, we provide a taxonomy of the SAS models and discuss the tradeo s between these models. 1 This material is based in part upon work supported by the Texas Advanced Technology Program und...
28 operational. An open area for research is in the implementation of the Resource Manager. This ... more 28 operational. An open area for research is in the implementation of the Resource Manager. This component i n teracts strongly not only with the operating system scheduler, but also with the disk subsystem and the graphics subsystem. We believe t h a t it is an open, but worthwhile challenge, to manage OS resource in the same way w e manage network resources in order to provide QoS guarantees in the end-system. We refer the reader to References 28, 2 2 , 2 3 , 2 4 as representative of work in this area. 10 Conclusion We h a ve described the design, implementation, and performance tuning of a native-mode ATM protocol stack. The transport layer provides three classes of service: reliable, guaranteed-service, and unreliable data transfer. An unusual feature is leaky-bucket policing at the transport layer for open-loop ow control. Our design is novel in that it is targeted to AAL5 and inexpensive P er-sonal Computers. We h a ve also tried to provide Quality of Service guarantees for re...
... The solution of these models is extremely ... seconds of processing are required for a model ... more ... The solution of these models is extremely ... seconds of processing are required for a model of a large contemporary mainframe computer system ... or specialized analytic techniques), is comparable to that achievable in other aspects of the performance evaluation process-workload ...
Proceedings of 1994 IEEE 10th International Conference on Data Engineering
ABSTRACT In this paper, we present a new replica control technique targeted at replicated systems... more ABSTRACT In this paper, we present a new replica control technique targeted at replicated systems in which write operations update a portion of the information in the data item rather than replacing it entirely. The existing protocols capable of supporting partial writes must either perform the writes on all accessible replicas of the data item, or always apply the writes to the same group (quorum set) of replicas. In the former case, the system incurs high message overhead. In the latter case, if any of the replicas in this group fail, additional replicas must be synchronously brought up-to-date during the write operation causing delay to the operation. Also, in both cases, the system loses the advantage of load sharing provided by replication. Our protocol avoids performing the write on all nodes while preserving load sharing and reducing greatly the risk of having to propagate updates synchronously. The paper describes the protocol, proves it correct, and presents a comparative performance s...
This paper concerns the parallel simulation of queueing network models (QNMs) using the conservat... more This paper concerns the parallel simulation of queueing network models (QNMs) using the conservative (Chandy-Misra) paradigm. Most empirical studies of conservative parallel simulation have used QNMs as benchmarks. For the most part, these studies concluded that the conservative paradigm is unsuitable for speeding up the simulation of QNMs, or that it is only suitable for simulating a very limited subclass of these models (e.g., those containing only FCFS servers). In this paper we argue that these are unnecessarily pessimistic conclusions. On the one hand, we show that the structure of some QNMs inherently limits the attainable simulation speedup. On the other hand, we show that QNMs without such limitations can be efficiently simulated using some recently introduced implementation techniques. We present an analytic method for determining an upper bound on speedup, and use this method to identify QNM structures that will exhibit poor simulation performance. We then survey a number ...
2008 IEEE Fourth International Conference on eScience, 2008
Scientific workflows have become an archetype to model in silico experiments in the Cloud by scie... more Scientific workflows have become an archetype to model in silico experiments in the Cloud by scientists. There is a class of workflows that are used to by "data valets" to prepare raw data from scientific instruments into a science-ready form for use by scientists. These share data-intensive traits with traditional scientific workflows, yet differ significantly, for example, in the required
ABSTRACT Structured programing is the set of design and implementation processes that yield well-... more ABSTRACT Structured programing is the set of design and implementation processes that yield well-structured programs. A well-structured program can be identified by the presence of five positive characteristics, and the absence of two negative characteristics.
Practice and Experience in Advanced Research Computing, 2021
CloudBank is a cloud access entity founded to enable the computer science research and education ... more CloudBank is a cloud access entity founded to enable the computer science research and education communities to harness the profound computational potential of public clouds. By delivering a set of managed services designed to alleviate common points of friction associated with cloud adoption, Cloudbank serves as an integrated service provider to the research and education community. These services include front-line help desk support, cloud solution consulting, training, account management, cost monitoring and optimization support, and automated billing. CloudBank has a multi-cloud pay-per-use billing model and aims to serve the spectrum of cloud users from novice to advanced.
because the underloaded hosts ask for chunks of work which involve number of threads which is jus... more because the underloaded hosts ask for chunks of work which involve number of threads which is just too small. In this case extreme load imbalance may occur: some hosts nish their part and stay completely idle, while others still have a lot of work that could otherwise be split. For instance, the system may reach a state in which 20 threads execute on one of the machines. The threads are divided into two groups of 10 threads each, where all the threads in each of the groups are communicating intensively with each other, but no communication exists between the groups. Clearly, one of these groups should be sent to one of the underloaded hosts, which may be prevented by the fact that 10 threads is greater than the overloaded threshold parameter of the load-balancing algorithm. Once again, this requires some smart mechanism which will adaptively identify and resolve this problem. We give here several additional parameters and issues that call for further research (or even simple tuning)...
The availability of 64-bit processors enables operating systems the use of the shared address spa... more The availability of 64-bit processors enables operating systems the use of the shared address space (SAS) paradigm where all processes execute in the same address space. This is in contrast to most existing operating systems that use the private address space (PAS) paradigm where each process views the entire space as dedicated to itself. The PAS paradigm yields high overhead for interprocess communication and sharing, whereas the SAS paradigm can reduce the cost of communication and provide simple abstractions to the application level to build variety of communication schemes. Furthermore, the use of the SAS paradigm results in increased system performance. The SAS paradigm can be classi ed into various models, which di er in the way the address space and protection domain are managed. In this paper, we provide a taxonomy of the SAS models and discuss the tradeo s between these models. 1 This material is based in part upon work supported by the Texas Advanced Technology Program und...
28 operational. An open area for research is in the implementation of the Resource Manager. This ... more 28 operational. An open area for research is in the implementation of the Resource Manager. This component i n teracts strongly not only with the operating system scheduler, but also with the disk subsystem and the graphics subsystem. We believe t h a t it is an open, but worthwhile challenge, to manage OS resource in the same way w e manage network resources in order to provide QoS guarantees in the end-system. We refer the reader to References 28, 2 2 , 2 3 , 2 4 as representative of work in this area. 10 Conclusion We h a ve described the design, implementation, and performance tuning of a native-mode ATM protocol stack. The transport layer provides three classes of service: reliable, guaranteed-service, and unreliable data transfer. An unusual feature is leaky-bucket policing at the transport layer for open-loop ow control. Our design is novel in that it is targeted to AAL5 and inexpensive P er-sonal Computers. We h a ve also tried to provide Quality of Service guarantees for re...
... The solution of these models is extremely ... seconds of processing are required for a model ... more ... The solution of these models is extremely ... seconds of processing are required for a model of a large contemporary mainframe computer system ... or specialized analytic techniques), is comparable to that achievable in other aspects of the performance evaluation process-workload ...
Proceedings of 1994 IEEE 10th International Conference on Data Engineering
ABSTRACT In this paper, we present a new replica control technique targeted at replicated systems... more ABSTRACT In this paper, we present a new replica control technique targeted at replicated systems in which write operations update a portion of the information in the data item rather than replacing it entirely. The existing protocols capable of supporting partial writes must either perform the writes on all accessible replicas of the data item, or always apply the writes to the same group (quorum set) of replicas. In the former case, the system incurs high message overhead. In the latter case, if any of the replicas in this group fail, additional replicas must be synchronously brought up-to-date during the write operation causing delay to the operation. Also, in both cases, the system loses the advantage of load sharing provided by replication. Our protocol avoids performing the write on all nodes while preserving load sharing and reducing greatly the risk of having to propagate updates synchronously. The paper describes the protocol, proves it correct, and presents a comparative performance s...
This paper concerns the parallel simulation of queueing network models (QNMs) using the conservat... more This paper concerns the parallel simulation of queueing network models (QNMs) using the conservative (Chandy-Misra) paradigm. Most empirical studies of conservative parallel simulation have used QNMs as benchmarks. For the most part, these studies concluded that the conservative paradigm is unsuitable for speeding up the simulation of QNMs, or that it is only suitable for simulating a very limited subclass of these models (e.g., those containing only FCFS servers). In this paper we argue that these are unnecessarily pessimistic conclusions. On the one hand, we show that the structure of some QNMs inherently limits the attainable simulation speedup. On the other hand, we show that QNMs without such limitations can be efficiently simulated using some recently introduced implementation techniques. We present an analytic method for determining an upper bound on speedup, and use this method to identify QNM structures that will exhibit poor simulation performance. We then survey a number ...
2008 IEEE Fourth International Conference on eScience, 2008
Scientific workflows have become an archetype to model in silico experiments in the Cloud by scie... more Scientific workflows have become an archetype to model in silico experiments in the Cloud by scientists. There is a class of workflows that are used to by "data valets" to prepare raw data from scientific instruments into a science-ready form for use by scientists. These share data-intensive traits with traditional scientific workflows, yet differ significantly, for example, in the required
ABSTRACT Structured programing is the set of design and implementation processes that yield well-... more ABSTRACT Structured programing is the set of design and implementation processes that yield well-structured programs. A well-structured program can be identified by the presence of five positive characteristics, and the absence of two negative characteristics.
Uploads
Papers by E. Lazowska