KR101395699B1 - Method for classifying task and method for improving performance of shared last level cache - Google Patents

Abstract

A task classification method according to an embodiment of the present invention includes: measuring a size of I / O traffic generated by a specific task; Comparing the maximum size of the shared last level cache (LLC) included in the system with the size of the I / O traffic; And classifying the specific task as a malicious task or a general task based on the comparison result. Thus, it is possible to accurately and quickly discriminate which task is likely to cause cache contamination in a NIMA system environment through a simple algorithm I can do it.

Description

[0001] METHOD FOR CLASSIFYING TASK AND METHOD FOR IMPROVING PERFORMANCE OF SHARED LAST LEVEL CACHE [0002]

The present invention relates to a task classification method and a method for improving performance of a shared last level cache using the method. More particularly, the present invention relates to a task classification method in a NUMA system and a method for improving performance of a shared last level cache using the same.

[0002] A non-uniform memory access (NUMA) system is a kind of multiprocessing system that processes a plurality of jobs in parallel. The system includes a plurality of numeration nodes, The CPU (central processing unit) in the system is basically a system that solves the performance conflict problem by making it accessible to a memory independent of the specific numeral node.

Each CPU under this system consists of a largely independent L1 and L2 cache memory and a shared-state LLC (Last Level Cache) memory.

In order to improve the overall performance of the Numa system, it is necessary to solve the problem of cache pollution of the shared last level cache included in the Numa system. Cache corruption means that data in an existing cache line is replaced with new data and prevents reuse of the cache. That is, there is a close relationship between how much the reuse probability of the shared last level cache can be increased and the overall performance of the Numa system.

I / O concentration tasks for multimedia video, downloading on web, processing files, etc. among tasks that occupy schedule computing resources and perform specific programs may cause serious cache pollution problem due to low cache reuse probability have.

In particular, if such an I / O concentration task is performed with other tasks, such as a memory concentration task, in a minimus system environment including a shared last level cache, the cache pollution also causes a cache efficiency of the memory concentration task to deteriorate.

Therefore, there is an active research on a scheduling method capable of reducing cache pollution by properly classifying tasks on a system having a shared cache such as Numa system.

One of the existing scheduling methods is an animal classification method. When each task is executed together in a core pair sharing a cache, the degree of mutual influence between tasks performed simultaneously is determined based on cache accessibility and cache miss rate I understand. Each task is classified into one of four types using the degree of mutual influence between tasks.

In addition, a scheduling method for classifying each task based on IPC (Instruction Per Cycle) or stack distance difference is also introduced.

As one of the existing scheduling methods, Japanese Laid-Open Patent Application No. 2011-059777 (entitled " Task Scheduling Method and Multi-core System ") discloses a scheduling method for executing as many tasks as possible, / RTI >
However, there is a continuing need for a task classification scheme that can improve the performance of the entire system by using the existing scheduling method and the differentiated scheduling method.

It is an object of the present invention to provide a task classification method capable of preventing cache pollution by classifying each task according to a classification standard differentiated from an existing one.

In addition, some embodiments of the present invention provide a method for enhancing the performance of an overall system as well as a shared cache by separately isolating tasks that may cause cache contamination utilizing a certain number of nodes of the Numa system It has its purpose.

According to a first aspect of the present invention, there is provided a method of classifying a task in a NUMA system according to the first aspect of the present invention, Measuring; Comparing the maximum size of the shared last level cache (LLC) included in the system with the size of the I / O traffic; And classifying the specific task as a malicious task or a general task based on the comparison result.

Also, a method for improving performance of a shared last level cache (LLC) using a task classification method in a NUMA system according to a second aspect of the present invention includes: Comparing the maximum size of the LLC (LLC) with the size of I / O traffic generated by a particular task; Classifying the specific task as a malicious task or a general task based on the comparison result; Changing a flag of the specific task to an ON state when the specific task is a malicious task; And moving a task in which the flag is in an ON state to a preset nummer node for the malicious task among a plurality of numeration nodes of the numeration system.

According to the task classification method in the Numa system, which is one of the task solutions of the present invention described above, the maximum size of the shared last level cache included in the numeration system is compared with the size of I / O traffic generated by a specific task By classifying a specific task into one of two types of tasks, it is possible to distinguish tasks that can cause cache pollution accurately and quickly through a simpler algorithm.

According to the method for improving the performance of the shared last level cache using one of the task resolution methods of the present invention and the task classification method, the specific task classified as the malicious task is moved to the specific nummer node of the numeration system to isolate , The reuse probability of the shared cache, and the performance of the overall system.

FIG. 1 and FIG. 2 are graphs for explaining the reason why a specific task is classified as one of two kinds of tasks according to the present invention;
3 is a block diagram schematically illustrating a structure of a numeration system for a task classification method according to an embodiment of the present invention and a method for improving performance of a shared last level cache using the method.
FIG. 4 is a flowchart illustrating a task classification method according to an embodiment of the present invention. FIG.
FIG. 5 is a flowchart for explaining a task classification method according to another embodiment of the present invention. FIG.
6 is a flowchart for explaining a task classification method according to another embodiment of the present invention;
FIG. 7 illustrates a process of comparing the maximum size of the shared last level cache with the size of I / O traffic in FIGS. 4 through 6. FIG.
8 is a graph showing a relationship between a weight and a cache miss rate for calculating a moving average value,
FIG. 9 is a flowchart illustrating a method for improving performance of a shared last level cache using a task classification method according to an embodiment of the present invention; FIG.
FIG. 10 is a flowchart illustrating a method for improving performance of a shared last level cache using a task classification method according to another embodiment of the present invention. FIG.
11 to 13 are graphs for evaluating the performance improvement of the shared last level cache by performing evaluation experiments.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings, which will be readily apparent to those skilled in the art. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. In order to clearly illustrate the present invention, parts not related to the description are omitted, and similar parts are denoted by like reference characters throughout the specification.

Throughout the specification, when a part is referred to as being "connected" to another part, it includes not only "directly connected" but also "electrically connected" with another part in between . Also, when an element is referred to as "comprising ", it means that it can include other elements as well, without departing from the other elements unless specifically stated otherwise.

A task classification method according to an embodiment of the present invention and a method for improving a performance of a shared last level cache using the same are characterized in that each task is classified into one of two types according to a new classification standard And the overall system performance can be improved by reducing the cache pollution through the task classification.

Before describing the present invention in earnest, the reason for classifying a specific task into any one of the two kinds of tasks will be described with reference to FIG. 1 and FIG.

I / O convergence tasks for multimedia video, downloading on the Web, processing of files and the like among tasks generally generate I / O traffic for a long period of time, and the I / O traffic of the big I / O task task.

FIG. 1 illustrates a case where a workload including a memory concentration task according to an exemplary embodiment and a workload including a I / O concentration task (PostMark or Grep) are performed together in one node, , The cache miss rate in the shared last level cache is compared. Since different shared caches are utilized for each node, it can be seen that performing both tasks on separate nodes, as shown in the graph of FIG. 1, results in a significantly lower cache miss rate. As a result, it can be seen that sorting and processing tasks that may cause cache pollution such as I / O concentration tasks can improve cache utilization efficiency.

FIG. 2 illustrates comparison of cache miss rates in a shared last level cache when two workloads (Tar) including an I / O concentration task according to an exemplary embodiment are performed together at one node and separately at another node This is a result. It can be seen that the cache miss rate is similar regardless of whether the I / O concentration tasks are classified as shown in the graph of FIG. It can also be seen that if both tasks are performed separately on different nodes, they still exhibit a significant cache miss rate. It can be seen from this that even if you carefully classify tasks that may cause cache pollution such as I / O intensive tasks, they do not have a significant impact on cache miss rates.

Accordingly, in the present invention, a simple algorithm for determining which kind of task corresponds to a particular task can be applied, so that a task that can cause cache contamination can be identified accurately and quickly.

3 is a block diagram briefly showing a structure of a numeration system for a method of classifying a task according to an embodiment of the present invention and a method for improving performance of a shared last level cache using the same.

The overall system 100 can be implemented through a kernel 10 that is the basis of software and includes a plurality of numeral nodes 20 and 30 and a storage 40, a physical main memory 50, a network 60 ), And the like. Each of the sleep nodes 20, 30 includes a CPU and a shared last level cache with one or more cores. C0, C2, C4, and C6 in the numeral node 20 basically access the memory in the numeral node 20 and perform various arithmetic operations. This is because the speed of accessing the memory in the numeration node 20 is faster than the speed of accessing the memory in the other numeration node 30.

In the present invention, the malicious identity recognition scheduler 11 provided by the kernel 10 monitors the I / O traffic generated by the task 12, and if the specific task performed through the CPU is the general task A, The malicious task B is moved to another numeral node 30 as indicated by the arrow in Fig.

Hereinafter, a task classification method according to each embodiment of the present invention will be described in detail with reference to FIG. 4 to FIG.

4 is a flowchart illustrating a task classification method according to an embodiment of the present invention.

First, when task classification according to an embodiment of the present invention is started, the size of I / O traffic generated by a specific task is measured (S410).

Next, the maximum size of the shared last level cache included in the numa system and the measured size of the I / O traffic are compared with each other (S420). As shown in FIG. 7, the maximum size of the shared last level cache is compared with the size of the I / O traffic.

Thereafter, a process of classifying a specific task as a malicious task or a general task is performed based on the comparison result (S430). That is, the size of the I / O traffic generated by the task and the maximum size of the shared cache in the case of a numeration system environment in which a plurality of numeration nodes having a shared last level cache exist are classified as tasks, .

In addition, a series of processes including the size measurement of the I / O traffic, size comparison, and determination of the type of a specific task can be performed by the malicious identity recognition scheduler 11 described above.

5 is a flowchart illustrating a task classification method according to another embodiment of the present invention.

First, when task classification according to another embodiment of the present invention starts, the size of I / O traffic is measured in real time using an internal kernel function for recording an I / O log file of a specific task (S510).

Typically, the Linux operating system writes I / O log files to the '/ proc' directory for each task when I / O traffic from various devices in the kernel occurs. I / O traffic can be generated by system calls such as' read () ',' write () ',' send () ',' receive Can be written to '/ proc / PID / io' via an internal kernel function such as 'do_io_accounting ()'.

Next, the maximum size of the shared last level cache included in the Numa system and the measured size of the I / O traffic are compared with each other (S520). As shown in FIG. 7, the maximum size of the shared last level cache is compared with the size of the I / O traffic.

If the size of the I / O traffic is larger than the maximum size of the shared last level cache, the specific task is identified as a malicious task (S530). If the size of the I / O traffic is smaller than the maximum size of the shared last level cache In the case of the same case, the specific task is distinguished as a general task (S540). That is, the size of the I / O traffic generated by the task and the maximum size of the shared cache in the case of a numeration system environment in which a plurality of numeration nodes having a shared last level cache exist are classified as tasks, .

6 is a flowchart illustrating a task classification method according to another embodiment of the present invention.

First, when task classification according to another embodiment of the present invention starts, I / O traffic generated by a specific task is acquired at a preset time interval (S610).

Subsequently, a variable a weight is given to the current I / O traffic and the I / O traffic before the predetermined time, and a moving average (MA) value is calculated (S620).

For example, through the I / O accounting scheme provided by the kernel, current I / O traffic using the 'rchar', 'wchar' variable of a particular task can be obtained periodically. The current I / O traffic can be reflected in the moving average value by Equation (1) below.

Here, the weight α is a variable value having a very close relationship with a classification for a specific task, and it is possible to determine how much the currently measured I / O traffic is reflected in the moving average, and to prevent frequent movement And can control I / O traffic fluctuations.

In other words, as the weight α increases, the current I / O traffic is reflected more in the moving average value than the I / O traffic that was before the preset time. In addition, the calculated moving average value is a value that can be matched with the size of the I / O traffic.

Thereafter, the maximum size of the shared last level cache included in the numa system and the size or moving average of measured I / O traffic are compared with each other (S630). As shown in FIG. 7, the maximum size of the shared last level cache is compared with the size of the I / O traffic.

If the size of the I / O traffic is larger than the maximum size of the shared last level cache, the specific task is identified as a malicious task (S640). If the size of the I / O traffic is smaller than the maximum size of the shared last level cache In the same case, the specific task is distinguished as a general task (S650). That is, using the size of the I / O traffic (or the moving average value) generated by the task and the maximum size of the shared cache in the case of a numeration system environment in which a plurality of numeration nodes having a shared last level cache exist, And each task is classified.

8 is a graph showing a relationship between a weight and a cache miss rate for calculating a moving average value.

As already described above, the weight α is a parameter closely related to the performance of the malicious recognition scheduler 11 described above with reference to FIG. As shown in FIG. 8, when the moving average value is calculated by assigning the weight α to 0.8, the lowest cache miss rate can be obtained.

By using the task classification method according to each of the embodiments of the present invention described above, it is possible to accurately and quickly identify which tasks can cause cache contamination under a minimized system environment through a simple algorithm. .

Meanwhile, a method for improving the performance of the shared last level cache (LLC) using the task classification method according to an embodiment of the present invention will be described with reference to FIG.

In the method of improving the performance of the shared last level cache (LLC) according to the present invention, the results classified by the task classification method according to each of the embodiments of the present invention described above can be utilized without any limit.

9 is a flowchart illustrating a method for improving performance of a shared last level cache using a task classification method according to an embodiment of the present invention.

First, when the method for improving the performance of the shared last level cache according to an embodiment of the present invention starts, the maximum size of the shared last level cache included in the NUMAR system is compared with the size of I / O traffic generated by a specific task S910). As shown in FIG. 7, it is possible to compare the maximum size of the shared last level cache with the size of the I / O traffic.

Thereafter, a process of classifying a specific task as a malicious task or a general task is performed based on the comparison result (S920). That is, the size of the I / O traffic generated by the task and the maximum size of the shared cache in the case of a numeration system environment in which a plurality of numeration nodes having a shared last level cache exist are classified as tasks, .

At this time, if the specific task is classified as a malicious task, the process of changing the flag of the specific task to the ON state is performed (S930). Changing the flag to the ON state may correspond to changing the value of the variable associated with the flag in the structure that is held for each task (or thread).

Subsequently, a process of moving a task in a flag ON state to a preset numon node for a malicious task among a plurality of numnail nodes of the numa system is performed (S940). Referring to FIG. 3, the flag of the specific task B classified as the malicious task is changed to the ON state, and the malicious task B is moved to the preset nummer node 30 for the malicious task. This is preferably done in a minimun system comprising a plurality of numeral nodes and the moved malicious task B can be isolated and processed separately from the general task A. [

In addition, a series of processes including the size measurement of the I / O traffic, size comparison, size determination of a specific task, flag change, and malicious task movement can be performed by the malicious identity recognition scheduler 11 described above.

10 is a flowchart illustrating a method for improving performance of a shared last level cache using a task classification method according to another embodiment of the present invention.

First, when the method for improving the performance of the shared last level cache according to another embodiment of the present invention is started, if the size of the I / O traffic generated by the specific task is larger than the maximum size of the shared last level cache included in the NIMA system, (S1010). As shown in FIG. 7, it is possible to compare the maximum size of the shared last level cache with the size of the I / O traffic.

If the size of the I / O traffic is larger than the maximum size of the shared last level cache, the specific task is identified as a malicious task (S1020). If the size of the I / O traffic is smaller than the maximum size of the shared last level cache In the case of the same case, the specific task is distinguished as a general task (S1050). That is, the size of the I / O traffic generated by the task and the maximum size of the shared cache in the case of a numeration system environment in which a plurality of numeration nodes having a shared last level cache exist are classified as tasks, .

At this time, if the specific task is classified as the malicious task (S1020), the process of changing the flag of the specific task to the ON state is performed (S1030). Changing the flag to the ON state may correspond to changing the value of the variable associated with the flag in the structure that is held for each task (or thread).

Subsequently, a process of moving a task in a flag ON state to a preset numon node for a malicious task among a plurality of numon nodes of the numar system is performed (S1040). Referring to FIG. 3, the flag of the specific task B classified as the malicious task is changed to the ON state, and the malicious task B is moved to the preset nummer node 30 for the malicious task.

In addition, when a specific task is classified as a general task (S1050), a task in which a flag is in an OFF state is processed in a numon node excluding a preset numon node according to a scheduling policy provided by an operating system in which the nummer system operates (S1060).

3, the flag of the specific task A classified as the general task is maintained in the OFF state and the normal task A is deleted from the numeral node 20 excluding the preset nummar node 30 Process according to the scheduling policy.

Through the method of improving the performance of the shared last level cache using the task classification method according to each of the above embodiments of the present invention, it is possible to accurately and quickly determine which task can cause cache pollution Can be distinguished.

In addition, there is an advantage that the reuse probability of the shared cache and the performance of the overall system can be improved by isolating and isolating a specific task classified as a malicious task to a specific nummer node of the numeration system.

The effects of the method for improving the performance of the shared last level cache using the task classification method according to each of the above embodiments of the present invention will be described with reference to FIG. 11 to FIG. 11 to 13 are graphs for evaluating the performance improvement of the shared last level cache by performing evaluation experiments.

The system used for the evaluation experiment is a 'Dell PowerEdge T610' with two nodes and has the specifications shown in Table 1 below.

Hardware Configuration Specifications L1 Data Cache 4 * 32KB, 8-way L1 Instruction Cache 4 * 32KB, 4-way L2 Cache 4 * 256KB, 8-way L3 shared Last Level Cache 4MB, 16-way Memory 2GB in each node

In addition, all the measurements were evaluated by the performance perfume tool "perf" based on Linux kernel using Hardware Performance Counter (HPC).

In addition, the evaluation experiment used various workloads including each task, where 'gobmk' is the artificial intelligence workload associated with gameplay and 'hmmer' is the statistical model of multiple sequential sorting. 'merge' is associated with a memory-intensive task and categorizes the 2MB matrix size with a recursive 2-way merge sorting algorithm. 'postmerge' and 'tar' are related to I / O highlighting tasks.

The graph on the left side of FIG. 11 shows results of performing 'postmark', 'gobmk', 'hmmer', and 'merge' at the same time, and the graph on the right shows 'tar', 'gobmk', 'hmmer', and 'merge' . It can be seen that the degree of decrease of the shared last level cache miss rate is shown in all the results from (1) to (6), and especially when the memory concentration task is distinguished from the I / O concentration task.

FIG. 12 shows a case where memory intensive workloads are executed on the same node as the I / O converged workloads, and (2) and (5) And ③ and ⑥ are cases where the present invention is applied. As can be seen in ③ and ⑥, we could reduce the miss rate of shared last level cache of 'hmmer' performed with 'postmark' by 19.1%. On average, we could reduce the shared cache miss rate by 20.7%.

FIG. 13 is a graph showing the execution time reduction rate. In the case of performing 'postmark' and 'hmmer' together, execution time was saved by 1.48%, and when performing 'merge' and 'tar' We could save execution time as much as. On average, we could reduce the shared cache miss rate by 0.87%.

It will be understood by those skilled in the art that the foregoing description of the present invention is for illustrative purposes only and that those of ordinary skill in the art can readily understand that various changes and modifications may be made without departing from the spirit or essential characteristics of the present invention. will be. It is therefore to be understood that the above-described embodiments are illustrative in all aspects and not restrictive. For example, each component described as a single entity may be distributed and implemented, and components described as being distributed may also be implemented in a combined form.

The scope of the present invention is defined by the appended claims rather than the detailed description and all changes or modifications derived from the meaning and scope of the claims and their equivalents are to be construed as being included within the scope of the present invention do.

10: Kernel 11: Malicious recognition scheduler
20, 30: Numa node 100: Numa system

Claims (8)

A method for classifying a task in a NUMA system, wherein the malicious recognition scheduler is provided by a kernel,
Measuring a size of I / O traffic generated by a specific task in the system;
Comparing the maximum size of the shared last level cache (LLC) included in the system with the size of the I / O traffic; And
And classifying the specific task as a malicious task or a general task based on a comparison result of the comparing step
Task classification method.
The method of claim 1, wherein measuring the size of the I / O traffic comprises:
The size of the I / O traffic is measured in real time using an internal kernel function for recording the I / O log file of the specific task
Task classification method.
2. The method of claim 1, wherein measuring the size of the I / O traffic comprises:
Acquiring I / O traffic generated by the specific task at predetermined time intervals; And
And calculating a moving average value by assigning a variable weight to the current I / O traffic and the I / O traffic before the predetermined time
Task classification method.
4. The method of claim 3, wherein as the weight increases, the weighting unit assigns the weight so that the current I / O traffic is more reflected in the moving average value than the I /
Task classification method.
2. The method according to claim 1, wherein in the step of classifying the specific task
And classifying the specific task as a malicious task when the size of the I / O traffic is larger than the maximum size of the shared last level cache.
A method for enhancing performance of a shared last level cache (LLC) using a task classification method in a NUMA system, the malicious recognition scheduler being provided by a kernel,
Comparing a maximum size of a shared last level cache (LLC) included in the system with an amount of I / O traffic generated by a particular task in the system;
Classifying the specific task as a malicious task or a general task based on a comparison result of the comparing step;
Changing a flag of the specific task to an ON state when the specific task is a malicious task; And
And moving a task in which the flag is in an ON state to a preset nummer node for the malicious task among a plurality of numeration nodes of the numeration system
How to improve performance of a shared last level cache.
7. The method according to claim 6, wherein in the step of classifying the specific task
When the size of the I / O traffic is larger than the maximum size of the shared last level cache, the specific task is classified as a malicious task
How to improve performance of a shared last level cache.
The method according to claim 6,
Processing a task in which the flag is in an OFF state in a numone node excluding the preset numeration node according to a scheduling policy provided by an operating system in which the numeration system operates,
How to improve performance of a shared last level cache.

Images