KR20050066133A - Processes monitoring system for cluster and method thereof - Google Patents

Abstract

The present invention relates to a system and a method for monitoring process information running on each compute node in a cluster environment. When the information to be monitored every minute increases as the size of the cluster increases, a hierarchical structure is provided to the monitoring system. This provides a system and method that can be applied to large clusters.

By applying the monitoring system and method of the present invention, detailed process information of a process running in each node of the cluster can be monitored, and by storing the monitoring information in a database, monitoring using a web interface or the like without having to log in to the system directly. It is possible to selectively monitor the process information desired by the user according to various criteria such as user and CPU occupancy.

Description

Process Monitoring System for Cluster and Method

The present invention relates to a system and method capable of monitoring process information running on each compute node in a clustered environment.

A cluster is a system used to connect an independent computer system such as a personal computer to a network device and use it for a special purpose. It is a solution for providing high performance computing and high availability computing. It is widely used.

In addition, if an upgrade or expansion is required, only a few nodes or networks can be replaced or added without changing the entire system, thereby providing flexible scalability. These advantages are increasing the number of cases where university labs, businesses, and laboratories that need to execute complex scientific computational problems are building clusters to meet their computational needs.

Because clusters are a collection of independent systems with separate hardware and operating systems, a single system image technology is important so that this is seen as a unified computing resource. A single system image makes a user feel like a large computer with a collection of multiple computers.

The single system image can be considered in various factors such as user policy, file system, CPU and memory resources, system monitoring, etc. In the present invention, it can be said that the single system image is provided for monitoring system processes. A process is a program running on an operating system such as UNIX. It is very important for a system administrator or a user to obtain information about a process currently running in system monitoring.

The purpose of monitoring the cluster is to monitor the abnormality of some of the constituent nodes constituting the cluster or to identify the load concentration on a specific node from the system administrator's point of view. It can also be used to predict the usage pattern of a system by continuously monitoring the overall utilization of the system over time.

Cluster monitoring, on the other hand, requires users to understand the overall utilization of the current system to determine whether the computations they want to perform can be performed, and to find out which nodes are focused on which nodes are under heavy load. To do work on fewer nodes. Then, by grasping the status of the job currently being executed, the purpose is to determine whether the job is being executed as originally intended.

Monitoring the processes running in a cluster is the most common way for users to track their progress. By monitoring the amount of system resources (CPU occupancy, memory usage, etc.) that processes are using, you can keep track of the progress of your work.

Developed by Berkeley University as an existing cluster monitoring system, Ganglia is good at monitoring the overall utilization of the system, but has the disadvantage of not monitoring each process running on the configuration node one by one. However, there is an advantage that it is easy to grasp the change of the system over time by continuously accumulating the monitoring information over time by using the Round Robin Database (RRD).

A tool for monitoring processes in a cluster is an SCMS package from Cassart University, Thailand, which includes not only monitoring but also other functions for cluster management. However, in the process monitoring, the result of simply executing the ps command of Unix is not familiar to the user, and it is inconvenient to use because there is no filter function to select only necessary processes.

The present invention provides a system and method for monitoring information of processes running on each component node of a cluster.

In addition, when the information to be monitored every moment as the size of the cluster increases, it provides a system and method that can be applied to a large cluster by having a hierarchical structure in the monitoring system.

1 shows a process monitoring system configuration for a cluster invented in the present invention. The monitoring system of the present invention includes a monitoring daemon, a monitoring information collector, a database, and an interface. The information collector, database, and interface constitute a monitoring node.

First, the monitoring daemon runs on all target nodes to monitor and collects information about processes running on each node and sends them to the network. In Linux system, process monitoring can be done by using / proc file system. Since information such as instantaneous CPU occupancy can be obtained by monitoring continuously with time difference, it keeps the information once monitored and update it continuously. Will go out.

Next, the information collector (collector) is executed in the monitoring node (monitoring node) in the cluster, collects information from the monitoring daemon, and the collected information is stored in a database. The information collector updates the database by collecting monitoring information from the monitoring daemon periodically or as needed.

The database is a place for storing process information collected in the cluster. The monitored information is stored using RDBMS (Relational Database Management System), enabling fast and free retrieval. In addition, by keeping the past monitoring information in the database, it can provide the user with the situation changed over time.

The interface provides various interfaces for the user to conveniently use the monitoring information collected in the database. The web interface provides process information through a web browser and can be easily monitored without access to the system without having to log in to the system, while the console interface allows you to view monitoring information using simple commands in the console. Provide the functionality to the user.

Next, a method of monitoring system processes using the process monitoring system for the cluster of the present invention will be described.

First, in the collecting step of monitoring information, the monitoring information is first collected by the monitoring daemon and delivered to the information collector. The time to collect the monitoring information is determined by the information collector, and the monitoring daemon responds to the information collector's request by collecting the information.

Methods of delivering monitoring information to information collectors include text-based message delivery in XML format and data in binary format.

The method of using text-based message in the form of XML is to collect collected monitoring information as an XML document and send it to the information collector. By extension of XML, monitoring items can be added easily and easily. While there is an advantage that there is no restriction on the implementation method of the ruler, the burden of generating and parsing an XML document is entailed, and the amount of data to be transmitted is increased because the text-based message is transmitted, and thus, the speed is slow.

On the other hand, the binary data transfer method is to transfer the data collected by the monitoring daemon to the information collector as it is language-dependent. As a result, parsing is not necessary. On the other hand, when adding monitoring items, there is a disadvantage that all monitoring daemons and information collectors must be modified at the same time. In addition, the implementation method is subject to the constraint that the monitoring daemon and information collector must be written using the same programming language.

In the present invention, process information is collected using a text message in XML format, and FIG. 2 shows an example of collected process information.

Next, in the request and response step of the monitoring information, the process information obtained from the monitoring daemon is delivered to the information collector. In this case, information about all processes may be delivered, but only the processes desired by the information collector may be delivered. For this purpose, the PID (process ID) list of the process for which status information is to be delivered to the monitoring daemon. In other words, between the monitoring daemon and the information collector consists of the protocol structure of request and response. By presenting the constraints of the monitoring information to be collected in the request message, it is possible to speed up the monitoring and reduce the data transmission volume by sending only necessary data to the information collector.

As a method of changing the monitoring cycle, there may be a method by a start time of a process, a method by a change in resource usage of a process, or a method directly specified by a user.

The method by the start time of the process is to increase the monitoring period as the total execution time of the process becomes longer. System processes are started during the system boot process and continue to run. These processes do not need to be monitored often, so this approach can reduce overall monitoring information.

The method by changing the process resource usage is to monitor the usage of resources such as CPU or memory and transmit the changed information only when there is a change in resource usage over a certain range.

On the other hand, in the case of a process that is always running, such as a system daemon process, and does not change significantly with time, the amount of monitoring information can be reduced by lengthening the monitoring period.

In another embodiment, the present invention provides a method for monitoring an extended cluster, in which a plurality of information collectors are provided so as not to reduce performance by limiting the number of nodes monitored by each information collector. This is because when the size of the cluster grows, the information transmitted to the information collector increases, which may reduce the performance of the overall monitoring system. In this case, only the target nodes to which the monitoring daemon is allocated are monitored, as shown in FIG. The information is transmitted to each information collector so that the information of the target nodes assigned to each monitoring daemon is stored separately in the monitoring node. Here, the monitoring node is composed of an information collector and a database, and exists in a number corresponding to each monitoring daemon and stores information of target nodes allocated to each monitoring daemon.

At this time, the monitoring interface node provides an interface for the user to access the monitoring information distributed and stored in each monitoring node in a consistent manner. The monitoring interface node receives a user's search condition and performs an SQL query on a distributed monitoring node to provide a result to the user.

The process monitoring system for the cluster of the present invention can easily monitor information on processes running on all nodes of the cluster in a consistent manner regardless of the size of the cluster.

In particular, process information can be searched by various search conditions, such as by user and resource usage rate. Since the database is used, it is possible to search process status changes according to time changes.

In addition, fast and accurate information can be obtained by reducing the load on the system for monitoring.

1 is a structural diagram of a process monitoring system of the present invention.

2 is the collected process information.

3 is a system architecture diagram for monitoring an extended cluster.

Claims (5)

A monitoring daemon for monitoring the process; An information collector for collecting information from the monitoring daemon; A database for storing information collected from the information collector; Process monitoring system comprising an interface for providing process information. The process monitoring system according to claim 1, wherein a protocol structure of a request and a response is provided between the monitoring daemon and the information collector. The process monitoring system of claim 1, wherein the information collector and the database form a monitoring node, and a request and response protocol structure is provided between the interface and the plurality of monitoring nodes. The information collected from the target node is first monitored by the monitoring daemon and is transmitted to the information collector. The monitoring daemon suggests the limitations of the monitoring information to be collected in the request message, so that only the process desired by the information collector is provided. Process monitoring method comprising the step of requesting and responding to the monitoring information sent to the collector. 5. The monitoring daemon of claim 4, wherein in the monitoring information collection step, the monitoring daemon monitors and transmits only information of the assigned target nodes to each information collector. In the request and response step of the monitoring information, the monitoring interface node monitors the monitoring daemon. And querying a monitoring node configured as the information collector.