Fault Tolerant Message Passing
Principle investigator: Nehal Desai CCS-1
Co-investigators: Dean Prichard CCS-1
�������������������� �������� Andrey Mirtchovski CCS-1
Recent trends in the use of clusters for large scale distributed scientific calculations have moved away from moderate size clusters of large SMPs, toward very large clusters of smaller SMPs. With larger numbers of nodes and more complex network topologies, the issue of reliability becomes paramount. The exponential dependence of the reliability on the size of the clusters leads to a rapid decrease in the mean time between failures of the cluster, will render application level reliability schemes such as checkpoint/restart ineffective.� In this study, we propose to implement a message passing system which utilizes transparent process mirroring to build a fault tolerant scientific computing system. Our goal is to increase the usefulness of clusters for scientific computing by greatly increasing their reliability. This work will leverage existing work in our group on building message passing systems for a number of operating systems.
This proposal will address the issue of increasing the reliability of commodity clusters for scientific computing.� With the increasing popularity of large scale Beowulf type clusters for scientific computing come new challenges.� In addition to the standard challenges of a well performing and scalable cluster is the issue of reliability.� Distributed scientific applications are unique among computer applications because they are extremely intolerant of system failures.� That is, they require all the resources of a computer (e.g. CPU, network, storage, etc.) at all times.� Therefore, in order for a scientific calculation to complete properly, system reliability over long periods of time is a must.� Before we begin any discussion of reliability, we need to define a few terms:
� Fault tolerance is the ability of parallel or distributed systems to recover from component failures without loss of performance.
� High availability (HA) systems provide users with almost uninterrupted services and provide reduced service during component failure.� The goal of high availability systems is minimal service interruption, not fault tolerance.
� Continuously available (CA) systems are capable of providing users with uninterrupted services and are only minimally affected by component or system failures.
In the context of scientific applications, reliability is equivalent to the level of service provided by a CA system.
Currently, most scientific applications rely on checkpointing to provide reliability, but as the size of clusters grows the effectiveness of the checkpoint-restart paradigm will diminish.� Because for large clusters, the mean time between failures (MTBF) of any component may approach the time required to write/save the data file required to restart the calculation.� The situation is easily illustrated by a simple calculation.� Consider the case of a computer node with a 10000 hour (~1.15 years) mean time between failures.� With 10000 nodes the probability of all the nodes being up for 1 hour is only 0.5, e.g. the mean time to failure for the full system is 1 hour.� Functionally, the probability of failure increases exponentially with the number of nodes.�
The MTBF is the most generous metric that can be used to estimate system reliability.� In reality, experiences with large-scale clusters such as ASCI Blue Mountain and ASCI Q show that even a temporary interruption of any hardware or software component of the cluster environment will likely cause a scientific application to terminate.� One of the ironies of large-scale clusters is that as the computing power of the cluster increases the ability to run large scale distributed scientific codes to completion actually decreases.
For upcoming tera-scale clusters, the need to incorporate reliability into the design of a cluster becomes imperative.� For current large-scale scientific computing clusters, such as the ASCI Q machine, the estimated MTBF for the full system is on the order of a few hours, compared to a roughly a day for previous systems such as ASCI Blue Mountain/Pacific.� If current trends continue, the next generation of large-scale clusters will have MTBF on the order of a few minutes.� For such clusters the checkpoint-restart methods in current use will be of little help.� Machines of this scale will require new ways of dealing with this inherent un-reliability.� This study proposes to develop techniques to increase the reliability and therefore the usability of large-scale clusters.�
Increasing the reliability of computer systems is not a new issue and has been addressed in a number of different contexts (telecommunications, storage, financial systems, air traffic control, etc.) using a plethora of different techniques [1,2,6].
Before we address techniques for increased reliability, we must presume a fault model.� Faults can be thought as occurring at various levels in a cluster.� In this study we assume faults occur at the operational or functional level.�� Operationally, there are two sources of unreliability in a computer system: the nodes (CPU) and the network.� Node faults occur for a number of reasons including hardware and software failures, and manifest themselves in a variety of ways from producing incorrect results to total failure.� The other source of unreliability is network faults, which result from a loss of networking hardware or software.� Network fault tolerance, given its current commercial importance, has been studied extensively resulting in a number of distinct methodologies .� In this study, we build and extend, previous work in message passing by implementing a fault-tolerant routing network.�
In addition to the source of the faults, fault models make assumptions about the way in which faults happen or are distributed.� In the current study, we assume a random fault model, in which components fail independently.�� This assumption greatly circumscribes the research, but allows for the introduction of various analytic modeling techniques.� In the later stages of this project we hope to remove this assumption, and incorporate more complex models into the fault tolerance framework.
One of the most promising methods for dealing with node faults is process mirroring.� In process mirroring, processes used in the calculation are replicated or mirrored.� The original and replicated processes are managed as a single unit in order to maintain identical computation and process state information.� Periodically, information is exchanged between the replicated processes, allowing a comparison of the calculation and the detection of faulted nodes.� The heuristics and algorithms used to determine which information is exchanged and how to arbitrate the different failure scenarios is key to the effectiveness process mirroring.� Process mirroring has been available for commercial purposes such as database and enterprise resource planning for many years.� However, to date no one has effectively applied this technique to distributed scientific computing.
Process units are of two types: simple and replicated.� Simple process units are single non-replicated processes.� Replicated units are multiple processes that are addressed as a single unit. In figure 1, processes 0, 1 and 3 are examples of simple processes while process 2 is an example of a replicated process.
Figure 1: Example of simple (left) and replicated (right) process units.
To understand the impact that process mirroring has on fault tolerance we need to examine some basic probability theory.� The probability of a single node failure as a function of time is given by the exponential distribution.
�������� p(X>t) = 1 - ebt
where b can be calculated from the MTBF� as:
b = -log(1/2)/MTBF
Considering node failures to be independent events, the probability of having any single node fail is:
q(t) = (1 - p(t))n
where n is the number of nodes.
With process mirroring the probability of any replicated process unit failing (both processes in a pair failing) is given by:
q(t) = (1 - p(t)2)n
where n is the number of replicated process units.
Figure 2 shows the example of MTBF = 10000 hrs, and n = 10000 process units.� The left hand panel shows the probability of success as a function of time without process pairs, while the right hand panel shows the same results with process pairs.� It is clear from the figure that replicating processes leads to a huge improvement in reliability.
Figure 2: Probability of job success as a function of job run time, left panel is without process replication, right panel is with process replication.
The goal of this study is to increase the usefulness of scientific computing clusters by increasing their reliability.� Because of the exponential dependence of the probability of failure on the number of nodes, increasing system reliability is the key to using commodity clusters to run tera-scale physics simulations.� This study will focus on using replicated processes and their extensions to improve overall machine reliability.� Increasing the reliability of clusters means users can spend more time doing science rather than marshaling the checkpoint-restart process.
We propose to implement a clustering infrastructure, which utilizes process mirroring and network resiliency to build a fault tolerant continuously available computer cluster.� This infrastructure will consist of several components including:
� The messaging layer which controls the flow of data between processes.� Since message passing is the primary mechanism to exchange information in distributed/parallel scientific computing, much of the reliability will be implemented at this level.�
� Autonomous network and node monitoring in order to determine current hardware and software status.
The messaging passing component will leverage our previous work in this area [3,5].� Much of the original messaging passing work in our group was driven by a re-examination of current models of scientific computation and questions as to whether those models will allow scientist to fully exploit the ever increasing power of commodity components.� One of the results of these explorations was the adoption of the Plan 9 operating system  as a prototyping/research environment.� The Plan 9 operating system is unique and we believe holds great potential for cluster research.� Plan 9 is appealing for number of reasons:
� Plan 9 is a distributed operating system with a global file system and novel security model.
� In Plan 9 all system devices are represented as files. Thus both local and remote resources can be easily controlled and accessed.� This allows for the rapid prototyping of new ideas.
� Support for commodity hardware and standard network protocols.
� Operating system support for checkpointing and process migration.
� Similar syntax to system V UNIX, allowing easy porting of applications between plan 9 and traditional Unix systems.
R & D approach and expected results.��
This phase of the study will involve adding transparent support for process pairs to our research messaging system.�� From a research perspective, the objective is not simply to implement process mirroring, but to develop to a general framework by:
1. Categorizing the different failure scenarios.
2. Determining which scenarios are most likely and should be addressed.
3. Determining the minimal (and sufficient) set of information that needs to be exchanged between mirrored processes to allow for detection of various failure scenarios.
4. Constructing algorithms/heuristics that use this information to avoid the faulty or dead nodes.
In addition there are a number of practical considerations, the framework (and its accompanying infrastructure) must address:� the availability of resources and the operational needs of users.� In some modes of operation users require only a high degree of certainty about the accuracy of the calculation and in others modes they require a high degrees of certainty of job completion.
In phase 1, all calculation� processes are mirrored.� Adding complete process mirroring to an existing distributed calculation results in an increase in the number of resources required to run a simulation.��� These additional resource requirements must be weighed against the need to successfully run a simulation (e.g. for mission critical applications like those in the ASCI program, the choice is obvious).� Figure 3 shows the additional nodes and network connections needed to implement process replication for a simple grid communication pattern.
This increase in resource usage has a large impact on the probability of a distributed calculation running to completion.� For example in the case described above, the addition of process replication increases the probability of the job continuing to run after 1 hour from 0.5 to greater than 0.99.
Figure 3: An example of the extra network connections and mirror nodes required for process replication.
In addition, to the initial fault tolerant infrastructure we will develop a test suite of representative application which can be used to demonstrate the effectiveness of the process replication and the fault tolerant message routing schemes.
While process replication greatly increases the fault-tolerance of a distributed computing system, the overhead of mirroring every process maybe too high for some applications.� In this phase of the study we will consider replication of only some of the processes.
If f is the fraction of nodes that have process pairs then the probability of success is given by:
q(t) = (1-p(t)2)nf * (1-p(t))n(1-f)
Figure 4 shows the probability of success as a function of the fraction of replicated pairs for run times of 1, 2, 4 and 8 hours.� One can see that the longer the run time, the more important it becomes to replicate all processes because of the high probability of failure for the un-replicated nodes.
Figure 4: Probability of job success as a function of the fraction of nodes that are replicated. Solid line is for a run time of 1 hour, dotted is for 2 hours, dashed is for 4 hours, and dot-dashed is for 8 hour run time.�
We also propose to use information obtained from system monitoring  and the history of the node behavior to estimate which nodes are more likely to fail in order to determine which processes will be mirrored.� As with Phase 1 we will demonstrate the software we develop using a test suite of applications on clusters in the ACL and document the results in a series of reports.
 R. Birman, Spinglass: Adaptive Probabilistic Tools for Advanced Networks http://www.cs.cornell.edu/Info/Projects/Spinglass/
 J. Bruck R. Cypher and C.-T. Ho., Fault-Tolerant Meshes of Hypercubes with Minimal Numbers of Spares. IEEE Transactions on Computers, 43, 9, 1089-1104. 1991.
 R. L. Graham, M W. Sulalski, L. D. Risinger, D. J. Daniel, N. N. Desai, M. Nochumsen, and L.-L. Chen, Performance analysis of the LA-MPI communication library. Los Alamos National Laboratory Technical Report LA-UR-02-1004.
 Ron Minnich and Karen Reid, Supermon: High performance monitoring for Linux clusters.
The Fifth Annual Linux Showcase and Conference, Oakland, CA.November 5-10, 2001.
 A. A. Mirtchovski, D. Prichard, and N. N. Desai, Using the Plan9 Operating System in Cluster Computing For Scientific Applications. Los Alamos National Laboratory Technical Report. Submitted to Cluster 2002.
 P. T. Murray, R. A. Fleming, P. D. Harry, and P. A. Vickers, Somersault: Enabling Fault-Tolerant Distributed Software Systems. Hewlett-Packard� Labs Technical Report. HP-98-81, 1998.
 R. Pike, D. Presotto, S. Dorward, B. Flandrena, K. Thompson, H. Trickey, and P. Winterbottom,� Plan9 from Bell Labs. Computing Systems, 8(3):221-254, 1995.
Ph.D��� Mechanical Engineering, North Carolina State University.� Work on multiphasic fluid mechanics.
Relevant Work Experience:
����������� Los Alamos National Lab (Oct 2000-Present)
Technical Staff Member, CCS-1.� Member of Cluster Research Team and� LA-MPI (Los Alamos Message Passing Interface) teams.
����������� IBM (June 1999-Oct. 2000)
Research Programmer.� Advanced Middleware Group.� Worked on several IBM products including:� Personal Communications Manager (PCOMM) and IBM Websphere
SGI (April 1998-June 1999)
����������� System Engineer for the ASCI BlueMountain System at Los Alamos
����������� Lab.� Part of the BlueMountain integration team.