KR101104999B1 - Load balancing method and system for metadata service - Google Patents

Abstract

PURPOSE: A load balancing method for a metadata service and system thereof are provided to supply a seamless service and to minimize the intervention of a manager in a cloud computing environment. CONSTITUTION: An MDS watcher periodically receives a heartbeat from an MDS(Metadata Server). An MDS controller calculates an MDS use value of the MDS value through the heartbeat. The MDS controller determines that the MDS having small MDS value is determined as a shadow MDS and copies a hotspot file to the shadow MDS. If a client inquires a hot spot file, a channel controller connects client to the shadow MDS. An update manager transmits the change information of a metadata lookup table to an MLTS(Metadata Lookup Table Server).

Description

Load balancing method and system for metadata service {LOAD BALANCING METHOD AND SYSTEM FOR METADATA SERVICE}

The present invention relates to a load balancing method and system for a metadata service, and more particularly, to a load balancing method and system for a metadata service using a master metadata server based on cloud computing.

The demand for cloud computing is increasing as cluster-based storage systems are increasingly demanding to take advantage of the already introduced storage using low-cost idle equipment without adding additional hardware.

In the field of cloud computing, virtualized storage technology is one of the important factors, and the combination of storage technology and virtualization technology based on existing cluster file systems is developing as a core technology of cloud computing and efficient I / O of file system. The share of metadata (MD) management for organizations is increasing. However, issues such as network traffic growth, bottlenecks and load balancing are still unresolved.

As described above, an object of the present invention is to provide a load balancing method and system for a metadata service which is a problem in cloud computing.

One aspect of the present invention is to place a master metadata server (MMDS) and a metadata lookup table server (MLTS) between a client and a plurality of metadata servers (MDS), and each MDS transmits a heartbeat only to the MMDS to allow meta The present invention relates to a method for performing load balancing for a data service.

In order to achieve the above object, a load balancing performing method according to the present invention includes a first step of receiving MMDS periodically receiving a heartbeat from an MDS; A second step of the MMDS using the heartbeat to calculate an MDS utilization value for each MDS; A third step of determining, by the MMDS, the MDS utilization value for each MDS, the MDS whose MDS utilization value is greater than or equal to the hotspot determination reference value as a hotspot MDS; If there is an MDS determined as a hotspot MDS in the third step, the MMDS checks the MDS utilization value for each MDS to determine the shadow MDS of the MDS having the lowest MDS utilization value among the MDSs whose MDS utilization value is less than the hotspot determination threshold. A fourth step; A fifth step of the MMDS copying the hotspot file existing in the hotspot MDS to the shadow MDS; When a hotspot file is queried by the client, the MMDS includes a sixth step of connecting the client to the shadow MDS, wherein the heartbeat indicates information indicating the state of the hardware of the MDS and the state of the metadata service latency of the MDS. Contains information that represents. Here, the information representing the state of the hardware of the MDS includes the maximum throughput of the CPU, the current throughput of the CPU, the maximum throughput of the RAM, and the maximum throughput of the RAM, and the information representing the state of the metadata service delay time of the MDS is MDS. Average latency of metadata in the queue and the number of update traffic in the MDS.

In addition, according to an embodiment of the present invention, the MLTS and the MDS each have a metadata lookup table, and after the sixth step, the MDS having changed the metadata lookup table sends a metadata lookup table change message to the MLTS. And a seventh step of transmitting, and an eighth step of MLTS transmitting the metadata lookup table change message to the MDS in which the change of the metadata lookup table has not occurred, and the change of the metadata lookup table does not occur. If there is more than one MDS, the order in which the MLTS transmits the metadata lookup table change message may be determined in consideration of the MDS utilization value.

Another aspect of the invention is a system for performing load balancing on metadata including a master metadata server (MMDS) and a metadata lookup table server (MLTS) disposed between a client and a plurality of metadata servers (MDS). It is about.

In order to achieve the above object, the MMDS includes an MDS watcher periodically receiving a heartbeat from the MDS and receiving metadata lookup table change occurrence information; The MDS utilization value for each MDS is calculated using the heartbeat, the MDS utilization value for each MDS is examined, and the MDS utilization value is determined as a hotspot MDS where the MDS utilization value is greater than or equal to the hotspot determination reference value, and the MDS utilization value is a hotspot. An MDS controller which determines a shadow MDS having the smallest MDS utilization value among MDSs that are less than a determination threshold value, and then copies the hotspot file existing in the hotspot MDS to the shadow MDS; A channel controller for connecting the client to the shadow MDS when the hotspot file is queried by the client; And an update manager for transmitting MDS utilization value and metadata lookup table change occurrence information for each MDS to MLTS, wherein the MLTS and MDS each have a metadata lookup table, and the heartbeat of the hardware of the MDS. Information indicating the status and information indicating the status of the metadata service delay time of the MDS. Here, the information representing the state of the hardware of the MDS includes the maximum throughput of the CPU, the current throughput of the CPU, the maximum throughput of the RAM, and the maximum throughput of the RAM, and the information representing the state of the metadata service delay time of the MDS is MDS. Average latency of metadata in the queue and the number of update traffic in the MDS.

In addition, according to an embodiment of the present invention, the MLTS receives a metadata lookup table change message from an MDS in which a metadata lookup table change occurs so that the metadata lookup table change message does not change the metadata lookup table. When there is more than one MDS transmitted to the MDS and no change of the metadata lookup table occurs, the order of transmitting the metadata lookup table change message may be determined in consideration of the MDS utilization value.

By the method and system according to the present invention, it is possible to minimize the intervention of the administrator in the cloud computing environment and to enable continuous service without interruption of service. In addition, low-cost storage servers can be utilized to maintain and manage large amounts of data. In particular, the MMDS specifically proposed in the present invention contributes to reducing network traffic, and the MLTS works in conjunction with the MMDS to enable efficient MLT update.

On the other hand, referring to Figures 5 to 7, through the performance evaluation according to the performance evaluation according to the number of MDS and the client, it can be seen that the performance by the method and system according to the present invention is significantly improved compared to the prior art.

1 is a schematic diagram of a load balancing system for a metadata service according to the present invention.
2 is a schematic diagram illustrating a configuration of a master metadata server (MMDS) included in a load balancing system for a metadata service according to the present invention.
3 is an algorithm for explaining load balancing according to the present invention.
4 is a flowchart illustrating a method for performing load balancing according to the present invention.
5 is a graph comparing metadata processing time according to a change in the number of MDSs according to the method and system according to the present invention and metadata processing time according to a change in the number of MDSs according to the related art.
6 is a graph comparing the metadata throughput according to the change of the MDS number in the case of the method and system according to the present invention and the metadata throughput according to the change in the MDS number in the case of the related art.
7 is a standard deviation of the number of metadata queries queued in the metadata server in the case of the method and system according to the present invention, and the standard number of metadata queries queued in the metadata server in the prior art. It is a figure comparing deviation.

1 is a schematic diagram of a load balancing system for a metadata service according to the present invention.

As shown in FIG. 1, in the present invention, a Master Metadata Server (MMDS) and a Metadata Look-up Table Server between a client and a plurality of Metadata Servers (MDSs). Load balancing for metadata services by deploying;

The MMDS periodically receives heartbeats from all MDSs. The heartbeat includes information indicating the state of the hardware of the MDS and information indicating the state of the metadata service delay time of the MDS. Therefore, the MMDS coordinates the load balancing of the MDSs through the information of the MDSs when a hot spot occurs.

This MMDS platform makes it possible to ensure scalability, flexibility, and stability for metadata services, and may utilize Xen schemes for virtualization in some embodiments.

The system according to the present invention utilizes queuing network theory to indicate the status of each MDS. It is assumed that the number of metadata queries in the system does not affect the arrival time and average service time of new metadata queries. If it is assumed that the arrival time of the new metadata query and the average service rate are constant, Equations 1 and 2 hold.

는 시스템으로 들어오는 평균 쿼리의 수를 나타낸다.here,

Is the average number of queries coming into the system.

는 시스템에 있는 쿼리의 평균 처리율을 나타낸다.here,

Represents the average throughput of queries in the system.

At this time, the load of the system can be obtained by the equation (3).

는 시스템의 부하를 나타낸다.here,

Is the load on the system.

Therefore, the probability that n is the number of queries in the system at steady state can be calculated by Equation 4.

은 시스템 내에 있는 쿼리의 수가 n일 확률을 나타내고,

는 수학식 5에 의해서 구할 수 있다. here,

Represents the probability that n is the number of queries in the system,

Can be obtained by equation (5).

In steady state, the probability of the number of queries in the system is n. It is assumed that the number of queries of metadata satisfies the relationship in equation (6).

은 시스템에 있는 평균 쿼리의 수를 나타낸다.here,

Is the average number of queries in the system.

)은 수학식 7에 의해서 구할 수 있다.The average time to wait in the queue for each MDS, that is, the average wait time for metadata in the MDS queue (

) Can be obtained by the equation (7).

와 같을 확률은 전확률의 법칙을 나타내는 수학식 8로부터 도출되는 수학식 9에서의 관계를 만족한다.Here we can find the probability that the time (T) the customer has to stay in the system is greater than t. Therefore, the probability that there are n queries existing in the system is as follows. percentage

Probability equal to satisfies the relationship in Equation 9 derived from Equation 8 representing the law of full probability.

확률은 수학식 10에 의해서 구할 수 있다.therefore,

Probability can be calculated by Equation (10).

), CPU의 현재 처리량(

), RAM의 최대 처리량(

), 및 RAM의 최대 처리량(

)을 포함한다. 또한, MDS의 메타데이터 서비스 지연 시간의 상태를 나타내는 정보는 MDS 큐에서 메타데이터의 평균 대기 시간(

) 및 MDS 내의 업데이트 트래픽의 수(

)를 포함할 수 있다.

에 따라서 CPU와 RAM의 활용도는 변화한다. 그러므로, 메터데이터 업데이트와 하트비트 메시지에 의해 영향을 받는

이 변할 것이다.Using the information indicating the state of the hardware of the MDS and the information indicating the state of the metadata service delay time of the MDS, the MDS utilization value can be obtained using Equation (11). In Equation 11, the hardware includes a CPU and RAM as an example, and in this case, the information representing the state of the hardware indicates the maximum throughput of the CPU (

), The current throughput of the CPU (

), Maximum throughput of RAM (

), And maximum throughput of RAM (

). In addition, the information indicative of the status of the metadata service delay time of the MDS may include the average latency of metadata in the MDS queue (

) And the number of update traffic in the MDS (

) May be included.

The utilization of the CPU and RAM changes accordingly. Therefore, affected by metadata updates and heartbeat messages

Will change.

는 MDS 활용도 값을 나타낸다.here,

Represents the MDS utilization value.

In the present invention, the MMDS periodically receives a heartbeat from the MDS to obtain an MMDS utilization value.

2 is a schematic diagram illustrating a configuration of a master metadata server (MMDS) included in a load balancing system for a metadata service according to the present invention.

Referring to FIG. 2, the MMDS seamlessly processes metadata and updates MLT through CC, UM, RC, MC, MW, XCM, XC, and RMSS components included in the MMDS as well as a relay role between the client and the MDS. To make it possible. The functions of each component are as follows.

The channel controller (CC) is responsible for connection management with the MDS to create and delete channels. For example, the channel controller may function to allow the MMDS to connect the client to the shadow MDS when the hotspot file is queried by the client.

The Update Manager (UM) receives the status of the MDS from the MC and reports the status of the current MDSs to the MLTS. For example, the update manager may transmit the MDS utilization value and the metadata lookup table change occurrence information for each MDS to the MLTS.

The relay controller (RC) establishes a direct connection between the client and the physical server.

The MDS Controller (MC) receives the status of the MDS from the MW, transmits to the desired controller and manages. For example, the MDS controller calculates the MDS utilization value for each MDS using the heartbeat, examines the MDS utilization value for each MDS, and determines an MDS whose MDS utilization value is greater than or equal to the hotspot determination threshold as a hotspot MDS. In addition, among the MDSs whose MDS utilization value is less than the hotspot determination threshold, the MDS having the smallest MDS utilization value may be determined as the shadow MDS, and then the hotspot file existing in the hotspot MDS may be copied to the shadow MDS.

The MDS Watcher (MW) observes the status and other information of the MDS and exchanges heartbeats. For example, the MDS watcher may periodically receive a heartbeat from the MDS and receive metadata lookup table change occurrence information.

3 is an algorithm for explaining load balancing according to the present invention.

Many connection requests from clients for a particular file in the filesystem occur frequently. In this case, if the file is replicated to more relaxed MDSs and can be distributed, clients can quickly process the file. Conventional methods update information about the file when a hotspot occurs and notify all MDSs and clients of this update. When the client approached, he requested to update the MLT version and used the updated information to find the shadow node to replace the MDS where the hot spot occurred and randomly select it. However, again, notifying all MDSs and clients of updates causes an increase in network traffic. In the present invention, when a hot spot occurs, the MMDS receives MDS utilization information from the MDS. The MMDS receiving this information can be processed according to the algorithm shown in FIG.

The MMDS periodically checks the MDS utilization value. If the MDS utilization value of a specific MDS is greater than or equal to α, the MDS is selected as a hotspot node. The simulation confirms the rapid increase in metadata processing delay from 10 MDS utilization values above α (eg, 0.7). However, if the number of MDSs, CPU, RAM, etc. change, the α value may change. When a particular MDS is selected as a hotspan node, it searches for shadow MDSs and duplicates the hotspot files that exist in the hotspan nose. The client does not know this process. Then, if the client tries to access the hotspot file, MMDS connects the client to another MDS with a replica file to resolve the hotspot. If the hotspot is released, the MMDS deletes the file on the shadow node.

4 is a flowchart illustrating a method of performing load balancing according to the present invention, and the algorithm as shown in FIG. 3 is reflected in the method of performing load balancing according to the present invention.

In the load balancing method according to the present invention, a step of periodically receiving a heartbeat from an MDS (S101), calculating an MDS utilization value (S103), checking an MDS utilization value (S105), and an MDS utilization value are reference values Determining the MDS as a hot spot MDS (S107), Determining the MDS having the smallest MDS utilization value as a shadow MDS (S109), Copying the hotspot file of the hotspot MDS to the shadow MDS (S111), and the hotspot file If the query is MMDS includes the step of connecting to the shadow MDS (S113).

In the method of performing load balancing according to the present invention, after the MDS utilization value is determined, the MDS having changed the metadata lookup table transmits a metadata lookup table change message to the MLTS, and the MLTS changes the metadata lookup table. The method may further include transmitting a message to an MDS in which a change of the metadata lookup table does not occur. When there are two or more MDSs in which the change of the metadata lookup table does not occur, the MLTS considers the MDS utilization value. You can specify the order in which the metadata lookup table change messages are sent. This will be described in detail below.

Table 1 is a table for explaining the metadata search table according to the present invention.

range Identification number MLT version 0-1000 One
20100225122425

1001-2000 2 2001-3000 3 3001-4000 4

In the present invention, as shown in Table 1, instead of using the result of hashing directly, the metadata lookup table (MLT) is used. Therefore, by modifying the MLT node without changing the hash function when the MDS node is added to or deleted from the cluster, or when renaming a file or changing a directory, the effect on the entire system, such as the modification of the hash function, is reduced.

라고 한다면 시스템의 총 지연시간은

만큼 연장되며 메타데이터 제공 서비스 효율은 감소한다.MLTS and all MDSs each have an MLT. The conventional DH (Dynamic Hashing) update method is updated when MDS nodes communicate with each other when an MDS node is added to or deleted from a cluster, or when there is a file name change or a directory change. This is not a problem for a few MDSs, but in a cloud computing environment, many nodes are added or removed, and file and directory renaming, deletion, and movement from many clients occurs. If multiple clients exchange update messages for each MDS in this manner, they will not only be inefficient, but will also incur a lot of network traffic between the MDSs. For example, if the MLT of five of the 10 MDSs needs to be updated, each of the five MDSs must send an update message to all nodes except themselves. A total of 45 update messages are then sent and processed, which reduces the effectiveness of the metadata providing service. This problem will become apparent as the number of MDSs increases. Therefore, increasing the network traffic caused by MLT update

The total latency of the system

As a result, the metadata providing service efficiency is reduced.

값이 발생하므로 시스템에 미치는 영향력은 감소하게 된다. 즉, MLT의 업데이트는 MMDS가

을 고려하여 우선순위에 따라 MLTS에게 여유있는 MDS의 MLT부터 업데이트 명령을 지시한다. The proposed update method does not inform each MDS of changes, and if there is a change of metadata in all MDSs, only the MLTS is notified so that the update message is updated considering CPU, RAM, and W.

As the value is generated, the impact on the system is reduced. In other words, the update of the MLT is

In consideration of this, the MLTS is instructed to update the MTS of the MDS of the MDS that is free in order of priority.

Since there is a version information area inside the MLT, the presence or absence of an update is checked using the version information. Nevertheless, it is possible to update periodically in case there is a node that cannot be updated due to continuous network traffic. If the client does not find the proper metadata in the MDS arriving through MLT due to hashing, the desired metadata can be obtained by updating the MLT in the MLTS from the current MDS without broadcasting to all MDSs.

The above description is merely illustrative of the technical details of the present invention, and those skilled in the art to which the present invention pertains may various modifications and changes without departing from the essential characteristics of the present invention. Therefore, the embodiments disclosed in the present invention are not intended to limit the technical idea of the present invention but to describe the present invention, and the scope of the technical idea of the present invention is not limited by these embodiments. The protection scope of the present invention should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present invention.

Claims (6)

Load balancing for metadata services by placing a master metadata server (MMDS) and a metadata lookup table server (MLTS) between the client and multiple metadata servers (MDS), and each MDS sends heartbeats only to the MMDS. As a method of performing
A first step in which the MMDS periodically receives a heartbeat from the MDS;
A second step of the MMDS using the heartbeat to calculate an MDS utilization value for each MDS;
A third step of determining, by the MMDS, the MDS utilization value for each MDS, the MDS whose MDS utilization value is greater than or equal to the hotspot determination reference value as a hotspot MDS;
If there is an MDS determined as a hotspot MDS in the third step, the MMDS checks the MDS utilization value for each MDS to determine the shadow MDS whose MDS utilization value is the smallest among the MDSs whose MDS utilization value is less than the hotspot determination threshold. A fourth step;
A fifth step of the MMDS copying the hotspot file existing in the hotspot MDS to the shadow MDS;
If the hotspot file is queried by the client, the MMDS includes a sixth step of connecting the client to the shadow MDS,
The heartbeat includes information indicating the state of the hardware of the MDS and information indicating the state of the metadata service delay time of the MDS. The method of claim 1,
Information indicating the state of the hardware of the MDS is the maximum throughput of the CPU (

), The current throughput of the CPU (

), Maximum throughput of RAM (

), And maximum throughput of RAM (

),
Information indicating the status of the metadata service latency of the MDS is based on the average latency of metadata in the MDS queue (

) And the number of update traffic in the MDS (

),
The MDS utilization value is calculated by Equation 1 below.
<Equation 1>

here,

Indicates the MDS utilization value. The method of claim 2,
MLTS and MDS each have a metadata lookup table,
After the sixth step, the MDS in which the change of the metadata lookup table has occurred, sends a metadata lookup table change message to the MLTS, and the MLTS sends the metadata lookup table change message to the metadata lookup table. And transmitting the metadata lookup table change message in consideration of the MDS utilization value when there is more than one MDS in which no change of the metadata lookup table occurs. How to perform load balancing for the metadata service, characterized in that for determining. A system for performing load balancing on metadata comprising a master metadata server (MMDS) and a metadata lookup table server (MLTS) disposed between a client and a plurality of metadata servers (MDS), the MMDS being:
An MDS watcher periodically receiving a heartbeat from the MDS and receiving metadata lookup table change occurrence information;
The MDS utilization value for each MDS is calculated using the heartbeat, the MDS utilization value for each MDS is examined, and an MDS whose MDS utilization value is greater than or equal to the hotspot determination reference value is determined as a hotspot MDS, and the MDS utilization value is a hotspot. An MDS controller which determines a shadow MDS having the smallest MDS utilization value among MDSs that are less than a determination threshold value, and then copies the hotspot file existing in the hotspot MDS to the shadow MDS;
A channel controller for connecting the client to the shadow MDS when the hotspot file is queried by the client; And
An update manager for transmitting MDS utilization value and metadata lookup table change occurrence information to each MLTS to MLTS;
MLTS and MDS each have a metadata lookup table,
The heartbeat is a system for performing load balancing for metadata services, characterized in that it includes information indicating the status of the hardware of the MDS and the metadata service delay time of the MDS. The method of claim 4, wherein
Information indicating the state of the hardware of the MDS is the maximum throughput of the CPU (

), The current throughput of the CPU (

), Maximum throughput of RAM (

), And maximum throughput of RAM (

),
Information indicating the status of the metadata service latency of the MDS is based on the average latency of metadata in the MDS queue (

) And the number of update traffic in the MDS (

),
The MDS utilization value is calculated by the following equation (1), the system for performing load balancing for the metadata service.
<Equation 1>

here,

Indicates the MDS utilization value. The method of claim 5, wherein
The MLTS receives the metadata lookup table change message from the MDS in which the metadata lookup table has been changed, transmits the metadata lookup table change message to the MDS in which the metadata lookup table has not been changed, and changes the metadata lookup table. And determining the order in which the metadata lookup table change message is transmitted in consideration of the MDS utilization value when two or more MDSs do not occur.

Images