CN104580476B - The method and apparatus for choosing node in a distributed system - Google Patents

Abstract

The present invention provides a method and device for selecting nodes in a distributed system, which helps improve the performance of the entire PrestoDB cluster by increasing the memory capacity of None nodes with relatively low workload and cost. The method includes: in the distributed system, using a specified part of the nodes as candidate None nodes; in the case that data pieces of multiple Source nodes need to be aggregated into one node, selecting among the candidate None nodes A node, and then gather the data slices of the multiple Source nodes to the selected node.

Description

Method and device for selecting nodes in distributed system

technical field

The invention relates to the field of computer technology, in particular to a method and device for selecting nodes in a distributed system.

Background technique

With the rise of big data, the amount of business data of Internet companies is increasing year by year. Therefore, major Internet companies are implementing big data technology internally and building data warehouses for core business systems. Currently, data warehouses are divided into two types: offline data warehouse and real-time data warehouse.

The representative product of offline data warehouse is hive. Since the underlying computing framework of this product is MapReduce, it is suitable for offline analysis and calculation of very large data sets, but it is not suitable for data analysis and calculation with high real-time requirements.

The representative product of the real-time data warehouse is PrestoDB. This product is developed by FaceBook and adopts the distributed data calculation and transmission mode of PipeLine. The analysis and calculation of big data can meet the requirements of real-time data analysis and calculation within 100ms-20m. requirements.

Since PrestoDB is a memory-based distributed computing framework, when performing data analysis and calculation, PrestoDB first divides the data to be analyzed and calculated into data slices and reads each data slice into the memory in the Source node of PrestoDB , and then aggregate the data in the memory of each Source node into a None node or multiple Fixed nodes through the network. The specific aggregation to the None node or the Fixed node is related to the type of aggregation function. For example: if the query contains order by statement, then all the results need to be sorted as a whole, so the data in the memory of each Source node needs to be aggregated into a None node, and then sorted as a whole; if the query contains a group by statement, then it is necessary to sort all The results are grouped, so the data in the memory of each Source node needs to be aggregated into multiple Fixed nodes for grouping.

At present, PrestoDB randomly selects a node from the entire cluster as the None node. The specific selection algorithm of various nodes in PrestoDB is shown in Figure 1. Figure 1 is a schematic diagram of the process of selecting nodes in the PrestoDB cluster according to the prior art. As shown in Figure 1, first determine the type of node to be selected. If you need to select a None node or a Fixed node, select it randomly in the cluster; if you need to select a Source node, first determine whether you need to use hardware-aware methods. Select according to data locality, otherwise randomly select multiple nodes as Source nodes. The hardware awareness here refers to the perception of the location of the data that needs to be processed, and the locality refers to the priority selection of the node where the data is located as the working node. Because if the assigned working node happens to be the node where the data to be processed is located, the time required for data transmission over the network can be reduced, and the time required for computing tasks can be reduced. Therefore, in some cases, the hardware perception method can be used to select nodes according to the principle of locality.

Therefore, it can be seen that if a node is selected as a None node, the requirement for its memory capacity is relatively large. In order to ensure the smooth progress of PrestoDB’s large-scale data analysis and calculation, it is necessary to upgrade the memory of all nodes in the cluster so that each node can meet the computing requirements when it is selected as a None node. The workload and cost of this upgrade are very high. bigger.

Contents of the invention

In view of this, the present invention provides a method and device for selecting nodes in a distributed system, which improves the performance of the entire PrestoDB cluster with relatively low workload and cost by only increasing the memory capacity of None nodes.

To achieve the above purpose, according to one aspect of the present invention, a method for selecting nodes in a distributed system is provided.

In the method for selecting nodes in a distributed system of the present invention, the distributed system is a PrestoDB cluster, and the method includes: in the distributed system, using a specified part of nodes as candidate None nodes; When the data pieces of the nodes are aggregated to one node, a node is selected among the candidate None nodes, and then the data pieces of the multiple Source nodes are aggregated to the selected node.

Optionally, after the specified part of the nodes are selected as the candidate None nodes, it also includes: judging whether the candidate None nodes are currently allowed under the condition that the data slices of multiple Source nodes need to be aggregated to multiple Fixed nodes As a candidate Fixed node, if yes, randomly select a plurality of nodes in the distributed system as Fixed nodes; otherwise, randomly select a plurality of nodes in the distributed system other than the candidate None nodes as Fixed nodes.

Optionally, after the specified part of the nodes are used as the candidate None nodes, it also includes: in the case of needing to save the fragmented data to the Source node, judging whether the candidate None nodes are currently allowed to be the candidate Source nodes , if so, randomly select multiple nodes in the distributed system as Source nodes, otherwise randomly select multiple nodes other than the candidate None nodes in the distributed system as Source nodes.

Optionally, the step of randomly selecting multiple nodes as Source nodes in the distributed system includes: in the case of currently adopting a hardware-aware manner, selecting multiple nodes as Source nodes in the distributed system according to the principle of locality node.

Optionally, the step of randomly selecting a plurality of nodes other than the candidate None nodes in the distributed system as Source nodes includes: in the case of currently using a hardware-aware manner, in the distributed system the In addition to the candidate None node, multiple nodes are selected as Source nodes according to the principle of locality.

According to another aspect of the present invention, a device for selecting a node in a distributed system is provided.

For the device for selecting nodes in a distributed system of the present invention, the distributed system is a PrestoDB cluster, and the device includes: a configuration module for recording a part of nodes designated as candidate None nodes in the distributed system; None A node selection module, configured to select a node among the candidate None nodes as a None node when the data pieces of multiple Source nodes need to be aggregated into one node.

Optionally, it also includes a Fixed node selection module, which is used to determine whether the candidate None node is currently allowed as a candidate Fixed node when the data pieces of multiple Source nodes need to be aggregated to multiple Fixed nodes, if , then randomly select multiple nodes in the distributed system as Fixed nodes, otherwise randomly select multiple nodes other than the candidate None nodes in the distributed system as Fixed nodes.

Optionally, it also includes a Source node selection module, which is used to determine whether the candidate None node is currently allowed as a candidate Source node when the fragmented data needs to be saved to the Source node, and if so, in the A plurality of nodes are randomly selected as Source nodes in the distributed system, otherwise, a plurality of nodes other than the candidate None nodes in the distributed system are randomly selected as Source nodes.

Optionally, the Source node selection module is further configured to select multiple nodes in the distributed system as Source nodes according to the principle of locality, in the case of currently adopting a hardware-aware manner.

Optionally, the Source node selection module is further configured to select multiple nodes in the distributed system as Source nodes according to the principle of locality, in the case of currently adopting a hardware-aware manner.

According to the technical solution of the present invention, a part of nodes are designated as candidate None nodes in the PrestoDB cluster, thereby limiting the selection of None nodes within a certain range, so that the memory upgrade and expansion of the nodes in this range can be performed to make them competent for computing Require. This method does not need to upgrade and expand the memory of all nodes in the entire PrestoDB cluster, so the workload of upgrading and expanding is relatively low, and it can improve the performance of the entire PrestoDB cluster.

Description of drawings

The accompanying drawings are used to better understand the present invention, and do not constitute improper limitations to the present invention. in:

Fig. 1 is a schematic diagram according to an embodiment of the present invention;

Fig. 2 is a schematic diagram of a method for selecting a node in a distributed system according to an embodiment of the present invention;

Fig. 3 is a schematic diagram of main modules of an apparatus for selecting a node in a distributed system according to an embodiment of the present invention.

Detailed ways

Exemplary embodiments of the present invention are described below in conjunction with the accompanying drawings, which include various details of the embodiments of the present invention to facilitate understanding, and they should be regarded as exemplary only. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the invention. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

In the solution of the embodiment of the present invention, a part of nodes in the PrestoDB cluster is designated in advance as candidate None nodes, and when it is necessary to select a None node, it is selected from this part of nodes. You can also set configuration items to configure whether to select None nodes from this part of nodes or randomly select None nodes in the PrestoDB cluster. When PrestoDB is started, the configuration item is parsed, and a list of corresponding IP-Port pairs is constructed according to the configuration information in the configuration item, and used when assigning None nodes. Its configuration specifications are, for example: None aggregation node = IP address 1: port 1; IP address 2: port 2. That is, two nodes whose IP addresses are address 1 and address 2 are designated as candidate None nodes, and the ports are port 1 and port 2 respectively. In the configuration item, it is also possible to configure whether to allow the above-mentioned candidate None node as a candidate Fixed node, and to configure whether to allow the above-mentioned candidate None node to be a candidate Source node. In this way, when a node needs to be selected, it can be performed according to the process shown in FIG. 2 . Fig. 2 is a schematic diagram of a method for selecting a node in a distributed system according to an embodiment of the present invention. This method can be executed by the Coordinator node in PrestoDB.

Step S21: Determine the type of node to be selected. In the case that the fragmented data needs to be saved to the Source node, the node to be selected is the Source node, and step S24 is entered. When data slices need to be aggregated, the type of node to be selected is determined according to the type of aggregation function. When a None node needs to be selected, go to step S22; when a Fixed node needs to be selected, go to step S23.

Step S22: Determine whether the None node is selected from the specified range. The judgment is made according to the above configuration items. If so, select a node from the candidate None nodes recorded in the configuration item as the None node (step S221), otherwise select a node randomly as the None node (step S222).

Step S23: Judging whether to allow the candidate None node to be the candidate Fixed node. If so, a plurality of nodes may be randomly selected as Fixed nodes (step S231 ), otherwise, a plurality of nodes other than candidate None nodes in the distributed system may be randomly selected as Fixed nodes.

Step S24: Determine whether hardware sensing is used, if yes, go to step S241, otherwise go to step S242.

Step S241: Judging whether to allow the candidate None node to be the candidate Source node. If so, multiple nodes can be selected as Source nodes according to the principle of locality (step S2411), otherwise multiple nodes can be selected as Source nodes according to the principle of locality in addition to the candidate None nodes in the distributed system (step S2412).

Step S242: Judging whether to allow the candidate None node to be the candidate Source node. If so, multiple nodes may be randomly selected as Source nodes (step S2421), otherwise multiple nodes other than the candidate None nodes in the distributed system may be randomly selected as Source nodes (step S2422).

Fig. 3 is a schematic diagram of main modules of an apparatus for selecting a node in a distributed system according to an embodiment of the present invention. As shown in FIG. 3 , the device 30 for selecting a node in a distributed system according to the embodiment of the present invention mainly includes a configuration module 31 and a None node selection module 32 . The configuration module 31 is used for recording a part of nodes designated as candidate None nodes in the distributed system. The None node selection module 32 is used for selecting a node among the candidate None nodes as a None node when the data slices of multiple Source nodes need to be aggregated into one node.

Apparatus 30 may also include a Fixed node selection module (not shown in the figure), which is used to determine whether the currently available None node is allowed to serve as Candidate Fixed nodes, if so, randomly select multiple nodes as Fixed nodes in the distributed system, otherwise randomly select multiple nodes as Fixed nodes other than the candidate None nodes in the distributed system.

The device 30 may also include a Source node selection module (not shown in the figure), which is used to determine whether the currently available None node is allowed as a candidate Source node when the fragmented data needs to be saved to the Source node, and if so, Then randomly select multiple nodes as Source nodes in the distributed system, otherwise randomly select multiple nodes as Source nodes other than the candidate None nodes in the distributed system.

The Source node selection module can also be used to select multiple nodes as Source nodes according to the principle of locality, in addition to the candidate None nodes in the distributed system when the hardware-aware method is currently used. The Source node selection module is also used to select multiple nodes as Source nodes in accordance with the principle of locality in addition to the candidate None nodes in the distributed system under the current hardware-aware mode.

According to the technical solution of the embodiment of the present invention, a part of nodes are designated as candidate None nodes in the PrestoDB cluster, thereby limiting the selection of None nodes within a certain range, so that the memory upgrade and capacity expansion of the nodes in this range can be performed, so that Competent computing requirements. This method does not need to upgrade and expand the memory of all nodes in the entire PrestoDB cluster, so the workload of upgrading and expanding is relatively low, and it can improve the performance of the entire PrestoDB cluster.

The above specific implementation methods do not constitute a limitation to the protection scope of the present invention. It should be apparent to those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (6)

1. a kind of method for choosing node in a distributed system, the distributed system are PrestoDB clusters, feature exists In this method includes：

In the distributed system, using specified a part of node as candidate None nodes；

In the case where needing the data slice by multiple Source nodes to converge to a node, in the None nodes of the candidate The data slice of the multiple Source nodes, is then converged to the node of selection by one node of middle selection；

Wherein, after using a part of node specified as candidate None nodes, further include：

In the case where needing the data slice by multiple Source nodes to converge to multiple Fixed nodes, whether judgement is currently permitted Perhaps the None nodes of the described candidate are as candidate Fixed nodes, if so, being randomly selected in the distributed system multiple Otherwise node randomly selects multiple sections as Fixed nodes except None nodes candidate described in the distributed system Point is used as Fixed nodes；

Wherein, after using a part of node specified as candidate None nodes, further include：

In the case where needing the data by fragment to be saved in Source nodes, the None for currently whether allowing the candidate is judged Node is as candidate Source nodes, if so, randomly selecting multiple nodes in the distributed system as Source Otherwise node randomly selects multiple nodes and is saved as Source except None nodes candidate described in the distributed system Point.

2. according to the method described in claim 1, making it is characterized in that, randomly selecting multiple nodes in the distributed system Include for the step of Source nodes：In the case of currently employed hardware perceptive mode, in the distributed system according to Locality principle chooses multiple nodes as Source nodes.

3. method according to claim 1 or 2, which is characterized in that the candidate None described in the distributed system Multiple nodes are randomly selected except node includes as the step of Source nodes：The currently employed hardware perceptive mode the case where Under, except the candidate None nodes described in the distributed system multiple node conducts are chosen according to locality principle Source nodes.

4. a kind of device for choosing node in a distributed system, the distributed system are PrestoDB clusters, feature exists In the device includes：

Configuration module, for recording appointed a part of node as candidate None nodes in the distributed system；

None node selecting modules, for the case where needing the data slice by multiple Source nodes to converge to a node Under, select a node as None nodes in the None nodes of the candidate；

Wherein, further include Fixed node selecting modules, for need the data slice by multiple Source nodes to converge to multiple In the case of Fixed nodes, the None nodes for currently whether allowing the candidate are judged as candidate Fixed nodes, if so, Multiple nodes are then randomly selected in the distributed system as Fixed nodes, otherwise described in the distributed system Multiple nodes are randomly selected as Fixed nodes except candidate None nodes；

Wherein, further include Source node selecting modules, in the feelings for needing the data by fragment to be saved in Source nodes Under condition, the None nodes for currently whether allowing the candidate are judged as candidate Source nodes, if so, in the distribution Multiple nodes are randomly selected in formula system as Source nodes, otherwise None candidate described in the distributed system is saved Multiple nodes are randomly selected as Source nodes except point.

5. device according to claim 4, which is characterized in that the Source nodes selecting module is additionally operable to adopt currently In the case of with hardware perceptive mode, multiple nodes are chosen as Source according to locality principle in the distributed system Node.

6. device according to claim 4 or 5, which is characterized in that the Source nodes selecting module is additionally operable to working as In the case of preceding use hardware perceptive mode, according to locality except the candidate None nodes described in the distributed system Principle chooses multiple nodes as Source nodes.