JP5809613B2 - Virtual machine tuning value calculation apparatus and virtual machine tuning value calculation method - Google Patents

Description

  The present invention relates to a virtual machine tuning value calculation device such as Java (registered trademark) VM (virtual machine) and a virtual machine tuning value calculation method.

  During garbage collection (GC) in a system using a virtual machine such as Java (registered trademark) VM, the system stops for a certain period of time due to copying and checking (marking) of objects. The mechanism of GC in Java (registered trademark) VM is disclosed in Non-Patent Documents 1 and 2 below. In general, in a system using a virtual machine, since various applications with different processing models are installed, it is difficult to predict (estimate) a program stop time by GC.

JavaVM Course for Tuning (Part 1), [online], [Search August 10, 2012], Internet <URL: http://www.atmarkit.co.jp/fjava/rensai3/javavm02/javavm02_1.html > JavaVM Course for Tuning (Part 2), [online], [Search August 10, 2012], Internet <URL: http://www.atmarkit.co.jp/fjava/rensai3/javavm02/javavm02_2.html >

  Here, in order to use a virtual machine in a system such as a communication node such as a SIP (Session Initiation Protocol) server, which performs a large amount of processing and requires low latency (reducing the delay time as much as possible) It is important to predict the program stop time by GC before development. However, as described above, in a system using a virtual machine, since various applications with different processing models are installed, it is difficult to predict (estimate) a stop time by GC.

  For this reason, the system developer needs to determine the tuning value of the virtual machine (the parameter value of the virtual machine that satisfies the target delay time (maximum allowable program stop time) and maximizes the throughput). We had to repeat the performance measurement and check the downtime. Here, what is particularly important for the tuning value is the size (NEW area size) allocated to the NEW area of the memory. Therefore, an object of the present invention is to provide means for solving the above-described problems and obtaining a NEW area size (optimum NEW area size) set in a virtual machine.

To solve the problems described above, the present invention provides a virtual machine tuning value calculating device for calculating the NEW region size of the memory to be set to the virtual machine used in the communication node, the virtual machine, NEW on the memory This is a virtual machine that adopts scavenge GC as the GC (garbage collection) method for the area and CMS (Concurrent Mark & Sweep) as the GC method for the OLD area on the memory. Applications used in the virtual machine are: A series of operations from the time when another terminal device accesses the virtual machine to the end of processing is managed as a call related to a plurality of control signals, and one or more long-lived call objects are generated as objects for the call And one or more short-lived control signals as an object for the control signal. A traffic model information indicating the number of calls generated per unit time (A1) and the number of control signals (A2) per call in the application, and one control signal. Per-call short-lived control signal object number (B1), long-lived call object number per call (B2), short-lived control signal object average hold time (B3), and short-lived control signal object size (B4) Application model information indicating the size (B5) of the long-lived call object, mark and copy time (C2) per object in the NEW area in the scavenge GC, and object copied to the OLD area in the CMS Remark time per one (C3) And an input of a target stop time (D), which is a target value of a program stop time at the time of GC in the virtual machine, is stored in a storage unit, and the traffic of the storage unit Using the model information, the application model information, the GC operation parameter information, and the target stop time, the NEW area size is calculated by the following equations (1) and (2). There are objects that are temporarily generated when a long-lived call object and a short-lived control signal object are generated, but these are also treated as short-lived control signal objects.
NEW area size = scavenge GC cycle T × number of calls per unit time (A1) × (number of long-lived call objects per call (B2) × long-lived call object size (B5) + control signal per call Number (A2) × number of short-lived control signal objects per control signal (B1) × size of short-lived control signal object (B4)) (1)
However,
Scavenge GC period T = {target stop time (D) − (number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) ) × short-lived control signal object average hold time (B3) × copy time per object in the NEW area (C2) in scavenge GC + number of calls generated per unit time (A1) × number of control signals per call ( A2) × number of short-lived control signal objects per control signal (B1) × average life of short-lived control signal objects (B3) × remark time per object copied to the OLD area in CMS (C3))} / (Number of calls per unit time (A1) x number of long-lived call objects per call (B2) x scavenge GC Mark and copy time per object in the NEW area (C2) + number of calls per unit time (A1) × number of long-lived call objects per call (B2) × objects copied to the OLD area in the CMS Remark time per one (C3)) ... Formula (2)

  As described above, the virtual machine tuning value calculation apparatus according to the present invention calculates (1) traffic model information, (2) application model information, and (3) GC of the virtual machine in order to calculate the NEW area size set in the virtual machine. Use operating parameter information. Here, in the conventional Web system, a large number of applications operate on one server (virtual machine), and the traffic conditions are different for each application. Therefore, (1) traffic model information and (2) application model information are It has been difficult to obtain a NEW area size (optimum NEW area size) by using it. However, if the computer in which the virtual machine is used is a communication node such as an SIP server, it is common that there is a single application that operates. Therefore, the inventor of the present invention pays attention to this point, and calculates (1) traffic model information and (2) application model information when calculating the NEW area size (optimum NEW area size) of the virtual machine used for the communication node. It came to devise a virtual machine tuning value calculation device to calculate using.

  Here, the application used in the virtual machine assumed in the present invention manages a series of operations from when another terminal device accesses the virtual machine to the end of processing as a call related to a plurality of control signals. To do. When the virtual machine processes one call, since the call exists in the memory for a relatively long time, one or more long-lived call objects are generated as objects for one call. On the other hand, when the virtual machine processes one control signal, since the control signal exists in the memory for a relatively short time, one or more short-lived control signal objects are generated as objects for one control signal.

  When this virtual machine performs GC, as described in Non-Patent Documents 1 and 2, the memory is divided into a NEW area and an OLD area to perform GC. Here, the scavenge GC method is adopted as the GC method for the NEW region, and the CMS method is adopted as the GC method for the OLD region. The program stop time during GC is the mark and copy time of an object from the NEW area to the OLD area in the scavenge GC + the remark time of the object copied to the OLD area in the CMS.

Here, in the present invention,
NEW area size (optimal NEW area size) = scavenge GC cycle T × number of calls per unit time (A1) × (number of long-lived call objects per call (B2) × long-lived call object size (B5) + Number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × size of short-lived control signal objects (B4)) (1)
It is said.

  In Equation (1), (number of calls generated per unit time (A1) × number of long-lived call objects per call (B2) × size of long-lived call objects (B5)) is a long-lived call in the NEW area per unit time. Indicates the total size occupied by the object.

  Also, in equation (1), (number of calls generated per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × short-lived control signal object The size (B4)) indicates an average value of the total size occupied by garbage of the short-lived control signal object and the short-lived control signal object being used in the NEW area per unit time.

  In other words, the NEW area size required for the virtual machine is the size occupied by garbage of the objects (short life control signal object in use and long life call object) existing in the NEW area during scavenge GC and the short life control signal object. Therefore, the NEW area size required for the virtual machine is the total size occupied by the objects (short-lived control signal object and long-lived call object) existing in the NEW area per scavenge GC period T × unit time.

Further, the scavenge GC period T used in the equation (1) is expressed by the following equation (2):
Scavenge GC period T = {target stop time (D) − (number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) ) X short-lived control signal object average hold time (B3) x mark and copy time per object in the NEW area in scavenge GC (C2) + number of calls generated per unit time (A1) x control signal per call Number (A2) × short-lived control signal object number per control signal (B1) × short-lived control signal object average holding time (B3) × remark time per object copied to the OLD area in CMS (C3)) } / (Number of calls per unit time (A1) x Number of long-lived call objects per call (B2) x Ska Mark and copy time per object in the NEW area in Benji GC (C2) + number of calls per unit time (A1) x number of long-lived call objects per call (B2) x copied to the OLD area in CMS Remark time per object (C3)) (2)
It is said. Note that the target stop time in Expression (2) is the maximum stop time allowed for the virtual machine by the system designer or the like.

  In Expression (2), (number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × short-lived control signal Average object holding time (B3) × mark and copy time per object in NEW area in scavenge GC (C2)) indicates the total mark and copy time of the short-lived control signal object in the NEW area in scavenge GC.

  Further, in Expression (2), (number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × short-lived control signal objects The average hold time (B3) × remark time (C3) per object copied to the OLD area in the CMS) indicates the total value of the remark times of the short-lived control signal objects copied to the OLD area in the CMS.

  Also, the number of calls per unit time (A1) × the number of long-lived call objects per call (B2) × the mark and copy time per object in the NEW area in the scavenge GC (C2) ) Indicates the total value of the mark and copy time of the long-lived call object in the NEW area in the scavenge GC.

  Further, in Expression (2), (number of calls per unit time (A1) × number of long-lived call objects per call (B2) × remark time per object copied to the OLD area in CMS (C3) ) Indicates the total remark time of the long-lived call object copied to the OLD area in the CMS.

  In other words, in the equation (2), the scavenge GC cycle T is expressed as follows: {target stop time (D)-(total value of the short life control signal object in the NEW area in the scavenge GC and the copy time + the short life copied in the OLD area in the CMS Total value of remark time for control signal object)} / (total value of mark and copy time of NEW area long-lived call object in scavenge GC + total value of remark time of long-lived call object copied to OLD area in CMS), , {Target stop time (D)-(Program stop time due to mark, copy and remark of short-lived control signal object)} / (Program stop due to mark, copy and remark of long-lived call object) ) To be. The scavenge GC period T is set to the target stop time (D)-(program stop time due to mark and copy and remark of short-lived control signal object) / (program stop time due to mark and copy and remark of long-lived call object). This is because it is assumed that the lifetime of the short-lived control signal object is shorter than the scavenge GC period T. That is, at the time of scavenging GC, not all short-lived control signal objects exist in the NEW area.

  Further, when the tuning parameter value calculation unit of the virtual machine tuning value calculation apparatus of the present invention calculates the NEW area size of each of the virtual machines in a system including a plurality of the virtual machines, the traffic of each of the virtual machines Using the model information, the application model information, the GC operation parameter information, and the target stop time of each of the virtual machines to which the target stop time of the entire system is distributed, the equation (1) and the equation (2) The NEW area size of each virtual machine is calculated.

  In this way, when the system is composed of a plurality of virtual machines, the NEW area size of each virtual machine that satisfies the target stop time requirement required for the entire system can be calculated.

The virtual machine tuning value calculating apparatus of the present invention, further receives an input of the target value of the operating rate of the virtual machine, and stored in the storage unit, the tuning parameter calculation unit sets the virtual machine When calculating the NEW area size, using the traffic model information, the application model information, the GC operation parameter information, the target stop time, and the target value of the operation rate in the storage unit, the following formula ( The range of the NEW area size is calculated from 3) to Expression (7).
Target stop time (D)> scavenge GC stop time t1 + remark stop time t2 Formula (3)
Scavenge GC cycle T> short-lived control signal object average holding time (B3) Equation (4)
Scavenge GC cycle T / (scavenge GC cycle T + scavenge GC stop time t1)> target value of operating rate (5)
However,
Scavenge GC stop time t1 = (number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × short-lived control signal object average Holding time (B3) + scavenge GC cycle T × number of calls per unit time (A1) × number of long-lived call objects per call (B2)) × mark and copy per object in NEW area in scavenge GC Time (C2) ... Formula (6)
Remark stop time t2 = (number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × short-lived control signal object average hold Time (B3) + scavenge GC period × number of calls per unit time (A1) × number of long-lived call objects per call (B2)) × remark time per object copied to OLD area in CMS ( C3) ... Formula (7)
When the NEW area size is set to be the left side, Expressions (3) to (5) are as shown in the following Expressions (8) to (11), and an optimum NEW area range is obtained from this. be able to.
NEW area size (optimum range) <scavenge GC cycle T × (number of calls per unit time (A1) × number of long-lived call objects per call (B2) × long-lived call object size (B5) + scavenge GC cycle T × number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × size of short-lived control signal objects (B4 )) ... Formula (8)
NEW area size (optimum range)> short-lived control signal object average hold time (B3) × (number of calls per unit time (A1) × number of long-lived call objects per call (B2) × long-lived call object Size (B5) + scavenge GC period T × number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × short-lived control Signal object size (B4)) (9)
NEW area size (optimal range)> target value of availability x number of courses per unit time (A1) x number of courses per unit time (A1) x (number of long-lived call objects per call) (B2) × long life call object size (B5) + number of control signals per call (A2) × number of short life control signal objects per control signal (B1) × short life control signal object size (B4)) X number of control signals per call (A2) x number of short-lived control signal objects per control signal (B1) x short-lived control signal object average hold time (B3) x per object in NEW area in scavenge GC Mark and copy time (C2)
/
(1—Target value of utilization rate—Target value of utilization rate × Number of course occurrences per unit time (A1) × Number of long-lived call objects per call (B2) × One object in NEW area in scavenge GC Mark and copy time (C2)) ... Formula (10)
However,
Scavenge GC period T = {target stop time (D) − (number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) ) X short-lived control signal object average hold time (B3) x mark and copy time per object in the NEW area in scavenge GC (C2) + number of calls generated per unit time (A1) x control signal per call Number (A2) × short-lived control signal object number per control signal (B1) × short-lived control signal object average holding time (B3) × remark time per object copied to the OLD area in CMS (C3)) } / (Number of calls per unit time (A1) x Number of long-lived call objects per call (B2) x Ska Mark and copy time per object in the NEW area in Benji GC (C2) + number of calls per unit time (A1) x number of long-lived call objects per call (B2) x copied to the OLD area in CMS Remark time per object (C3)) (11)

  By doing in this way, the virtual machine tuning value calculation apparatus, in addition to the program stop time of the virtual machine being less than the target stop time, the scavenge GC cycle is longer than the short-lived control signal object average hold time, and Considering that the operating rate of the virtual machine exceeds the target value, an appropriate NEW area size range set for the virtual machine can be obtained.

  Note that (scavenge GC stop time t1 + remark stop time t2) shown in Expression (3) indicates a program stop time during GC in the virtual machine. In addition, (scavenge GC cycle T / (scavenge GC cycle T + scavenge GC stop time t1)) in equation (5) indicates the operating rate of the virtual machine.

  Also, the number of calls generated per unit time (A1) × the number of control signals per call (A2) × the number of short-lived control signal objects per control signal (B1) in Expression (6) and Expression (7) X Short-lived control signal object average holding time (B3)) indicates an average value of the number of short-lived control signal objects that survive in the NEW area during scavenge GC.

  Expression (6) and Expression (7) (scavenge GC period T × number of calls per unit time (A1) × number of long-lived call objects per call (B2)) occur during the scavenge GC period. Indicates the number of long-lived call objects. That is, it indicates the number of long-lived call objects that survive in the NEW area during scavenging GC.

Regarding the remark stop time, the number of objects copied from the NEW area to the OLD area by the scavenge GC (that is, the number of live objects that live in the NEW area when the scavenge GC period is reached) × the objects copied to the OLD area It can be thought of as the remark time per one. Therefore, the remark stop time is as shown in Expression (12),
Remark stop time t2 = (number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × short-lived control signal object average hold Time (B3) + scavenge GC period × number of calls per unit time (A1) × number of long-lived call objects per call (B2)) × remark time per object copied to OLD area in CMS ( C3) ... Formula (12)
It is said.

  According to the present invention, a NEW area size (optimum NEW area size) set in a virtual machine used for a communication node can be easily obtained.

It is a functional block diagram of the virtual machine tuning value calculation apparatus of this Embodiment. It is a figure for demonstrating the traffic model information and application model information of this Embodiment. It is a figure for demonstrating GC in a virtual machine. It is a figure for demonstrating the calculation process of the tuning parameter value (optimal NEW area | region size) in the system which consists of a some server. It is the figure which illustrated the operation rate of the virtual machine with respect to the NEW area size of a virtual machine. It is a figure which illustrates the range of the optimal value of the value of the number of short-lived control signal objects x the size of the short-lived control signal object, and the number of long-lived call objects x the size of the long-lived call object.

  Hereinafter, an embodiment for implementing a virtual machine tuning value calculation apparatus of the present invention will be described with reference to the drawings. First, the configuration of the virtual machine tuning value calculation apparatus 10 according to the present embodiment will be described with reference to FIG.

<Configuration>
The virtual machine tuning value calculation apparatus 10 includes application model information (described later), traffic model information (described later), GC (garbage collection) operation parameter information (described later), and a target value (target) of the program stop time during GC of this virtual machine. When an input of stop time) is received, a NEW area size (optimum NEW area size) set in the virtual machine is calculated and output based on the information and the target value. The function of the virtual machine tuning value calculation device 10 is divided into an input unit 11, a tuning parameter value calculation unit 12 and an output unit 13.

  The input unit 11 and the output unit 13 are composed of input / output interfaces. The tuning parameter value calculation unit 12 is realized by a CPU (Central Processing Unit, not shown) included in the virtual machine tuning value calculation apparatus 10 developing and executing a program in a main memory (not shown). This program is stored in a predetermined area of an auxiliary storage unit (not shown) included in the virtual machine tuning value calculation apparatus 10. The tuning parameter value calculation unit 12 may be realized by a dedicated circuit.

  The input unit 11 includes an application model input unit 111 that receives input of application model information, a traffic model input unit 112 that receives input of traffic model information, a GC operation parameter input unit 113 that receives input of GC operation parameters, and an input of a target stop time The target stop time input unit 114 is provided. The output unit 13 includes a tuning parameter value output unit 131 that outputs a NEW area size set in the virtual machine. The tuning parameter value calculation unit 12 calculates a NEW area size set in the virtual machine by a tuning parameter value calculation algorithm, and outputs the calculated NEW area size to an external device or the like via the tuning parameter value output unit 131.

  Here, the virtual machine is, for example, a Java (registered trademark) VM (virtual machine) and is used for a communication node such as a SIP (Session Initiation Protocol) server. An application (application program) used in this virtual machine is a call involving a plurality of control signals, which is a series of operations from when another terminal device or the like accesses this virtual machine until the processing is completed. Shall be managed.

  For example, consider the case where the application is a telephone application. In this case, one telephone call is one call, and control signals such as “connection start” and “connection end” are related to this call. With such an application, the virtual machine generates one or more objects for the call and one or more objects for the control signal. Here, since processing for one call exists in the memory used by the virtual machine for a relatively long time, an object generated for one call is a long-lived call object. On the other hand, one control signal is generated for one control signal because the processing is completed in a relatively short time (predetermined time) and deleted from the memory even if an object is generated for the processing. Let the object be a short-lived control signal object.

  The GC of the virtual machine is performed in each of the NEW area and the OLD area of the memory, and there is a method in each area. Here, the scavenge GC method is adopted as the GC method of the NEW area, and the GC of the OLD area is determined. A case where the CMS method is adopted as a method will be described. A live object generated by a Java (registered trademark) VM is marked when a predetermined scavenge GC cycle is reached, and is copied from the NEW area to the OLD area. Then, the area reserved for the object in the NEW area is released. Here, the object copied to the OLD area is checked (remarked). The time for marking and copying the NEW area and the time for remarking the object copied to the OLD area are the program stop time by the GC of the virtual machine.

  The traffic model (traffic model information), application model (application model information), GC operation parameter information, and target stop time handled in the present embodiment will be described.

<Traffic model information>
The traffic model information includes the following information related to the application used by the virtual machine (see FIG. 2A).
-Number of calls per unit time (A1)
-Number of control signals per call (per call) (A2)
By using such traffic model information, the tuning parameter value calculation unit 12 can, for example, generate the number of control signals per unit time in the NEW area (number of calls generated per unit time (A1) × per call). The number of control signals (A2)) can be obtained.

<Application model information>
The application model information includes the following information related to the application used by the virtual machine (see FIG. 2B).
・ Number of short-lived control signal objects per control signal (per control signal) (B1)
・ Number of long-lived call objects per call (per call) (B2)
・ Short-life control signal object average holding time (B3)
・ Size of short-lived control signal object (B4)
・ Size of long-lived call object (B5)

<GC operation parameter information>
Further, the GC operation parameter information includes the following information used when performing GC in the virtual machine.
-Mark and copy time per object in the NEW area in the scavenge GC (C2)
-Remark time per object copied to OLD area in CMS (C3)

<Target stop time>
The target stop time (D) is a program stop time during GC that is allowed in the virtual machine.

  The tuning parameter value calculation unit 12 calculates a NEW area size set in the virtual machine by the virtual machine tuning value calculation algorithm. This virtual machine tuning value calculation algorithm uses the above-described traffic model (traffic model information), application model (application model information), GC operation parameter information, and target stop time (D). This is an algorithm for calculating the NEW area size of the virtual machine according to (2).

NEW area size = scavenge GC cycle T × (number of calls generated per unit time (A1) × number of long-lived call objects per call (B2) × size of long-lived call objects (B5) + number of calls generated per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × short-lived control signal object average hold time (B3) × short-lived control signal object size (B4) )) ... Formula (1)
However,
Scavenge GC period T = {target stop time (D) − (number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) ) X short-lived control signal object average hold time (B3) x mark and copy time per object in the NEW area in scavenge GC (C2) + number of calls generated per unit time (A1) x control signal per call Number (A2) × short-lived control signal object number per control signal (B1) × short-lived control signal object average holding time (B3) × remark time per object copied to the OLD area in CMS (C3)) } / (Number of calls per unit time (A1) x Number of long-lived call objects per call (B2) x Ska Mark and copy time per object in the NEW area in Benji GC (C2) + number of calls per unit time (A1) x number of long-lived call objects per call (B2) x copied to the OLD area in CMS Remark time per object (C3)) (2)

  The expressions (1) and (2) will be described with reference to FIGS. 2 (a), (b) and FIG. 3 indicates the size occupied by objects (live objects and non-live objects) in the NEW area size. In particular, the bold line in the graph 201 indicates the size occupied by the live object in the NEW area size. The graph 202 shows the size occupied by objects (live objects and non-live objects) in the OLD area size. This live object is an object called a root set, in which the reference is traced from the object that is the basis of the reference and connected by the reference. The non-live object is an object other than the live object among the objects on the memory. In GC, the memory area of this non-live object is released. The live object is an object that is connected by reference when the reference is traced from an object that is a reference source called a root set. The non-live object is an object other than the live object among the objects on the memory. In GC, the memory area of this non-live object is released.

  As described above, the virtual machine performs scavenging GC when an object generated on the NEW area (including a short-lived control signal object, a long-lived call object, and garbage) reaches the NEW area size, that is, when the scavenging GC period T is reached. . As a result, the live object in the NEW area is marked, copied to the OLD area, and the area reserved for the object in the NEW area is released. For example, as shown in FIG. 3, the virtual machine copies a live object among the short-lived control signal object and long-lived call object in the NEW area to the OLD area. Then, in the NEW area, the area allocated to the abandoned object (non-live object) and the copied live object is released, and the size occupied by the object in the NEW area size becomes 0 (see graph 201). . Thereby, an area for accepting a new object is secured in the NEW area. In addition, since it is assumed that this scavenge GC cycle T is longer than the short-lived control signal object average holding time (average survival time), the short-lived control signal object is not abandoned object (non-free) before reaching the scavenge GC cycle T. Live object). Further, the virtual machine marks the live object in the OLD area, and then remarks the live object copied from the NEW area. Then, sweeping of objects (abandoned objects) other than the object (live object) marked and remarked in the OLD area is performed. That is, the area allocated to the abandoned object is released. As a result, the size occupied by the object in the OLD area size decreases (see graph 202). Thereby, an area for accepting a new object is secured in the OLD area. It should be noted that the scavenge GC stop time t1 shown in FIG. 3 (the mark of the object in the NEW area and the copy time to the OLD area by the scavenge GC) and the remark stop time t2 (the live object group copied from the NEW area to the OLD area by the CMS) The total value of the remark time) is the program stop time during GC in the virtual machine.

  As described above, the tuning parameter value calculation unit 12 obtains the NEW region size (optimum NEW region size) by the following equations (1) and (2).

  NEW area size = scavenge GC cycle T × number of calls per unit time (A1) × (number of long-lived call objects per call (B2) × long-lived call object size (B5) + control signal per call Number (A2) × number of short-lived control signal objects per control signal (B1) × size of short-lived control signal object (B4)) (1)

  In Equation (1), (number of calls generated per unit time (A1) × number of long-lived call objects per call (B2) × size of long-lived call objects (B5)) is a long-lived call in the NEW area per unit time. Indicates the total size occupied by the object.

  Also, in equation (1), (number of calls generated per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × short-lived control signal object Average hold time (B3) × size of short-lived control signal object (B4)) indicates an average value of the total size occupied by the short-lived control signal object in the NEW area per unit time.

  In other words, the NEW area size required for the virtual machine only needs to be equal to the number of objects (short life control signal object and long life call object) existing in the NEW area during scavenging GC. , Scavenge GC period x total size occupied by objects (short-life control signal object and long-lived call object) existing in the NEW area per unit time.

  Further, in equation (1), the average value of the total size occupied by the short-lived control signal object in the NEW area per unit time is expressed as (number of calls generated per unit time (A1) × number of control signals per call (A2)) X number of short-lived control signal objects per control signal (B1) x short-lived control signal object average holding time (B3) x short-lived control signal object size (B4)) This is because it is assumed that it is shorter than the scavenge GC cycle T. That is, at the time of scavenging GC, not all short-lived control signal objects exist in the NEW area.

Further, the scavenge GC period T used in the equation (1) is expressed by the following equation (2):
Scavenge GC period T = {target stop time (D) − (number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) ) X short-lived control signal object average hold time (B3) x mark and copy time per object in the NEW area in scavenge GC (C2) + number of calls generated per unit time (A1) x control signal per call Number (A2) × short-lived control signal object number per control signal (B1) × short-lived control signal object average holding time (B3) × remark time per object copied to the OLD area in CMS (C3)) } / (Number of calls per unit time (A1) x Number of long-lived call objects per call (B2) x Ska Mark and copy time per object in the NEW area in Benji GC (C2) + number of calls per unit time (A1) x number of long-lived call objects per call (B2) x copied to the OLD area in CMS Remark time per object (C3)) (2)
It is said. The target stop time (D) in equation (2) is the maximum program stop time allowed for the virtual machine by the system designer or the like.

  Also, in equation (2), (number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × short-lived control signal object Average hold time (B3) × mark and copy time per object in NEW area in scavenge GC (C2)) indicates the total value of the mark and copy time of the short-lived control signal object in the NEW area in scavenge GC.

  Further, in Expression (2), (number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × short-lived control signal objects The average hold time (B3) × remark time (C3) per object copied to the OLD area in the CMS) indicates the total value of the remark times of the short-lived control signal objects copied to the OLD area in the CMS.

  Also, the number of calls per unit time (A1) × the number of long-lived call objects per call (B2) × the mark and copy time per object in the NEW area in the scavenge GC (C2) ) Indicates the total value of the mark and copy time of the long-lived call object in the NEW area in the scavenge GC.

  Further, in Expression (2), (number of calls per unit time (A1) × number of long-lived call objects per call (B2) × remark time per object copied to the OLD area in CMS (C3) ) Indicates the total remark time of the long-lived call object copied to the OLD area in the CMS.

  In other words, in the equation (2), the scavenge GC cycle T is expressed as follows: {target stop time (D)-(total value of the short life control signal object in the NEW area in the scavenge GC and the copy time + the short life copied in the OLD area in the CMS Total value of remark time for control signal object)} / (total value of mark and copy time of NEW area long-lived call object in scavenge GC + total value of remark time of long-lived call object copied to OLD area in CMS), , {Target stop time (D)-(Program stop time due to mark, copy and remark of short-lived control signal object)} / (Program stop due to mark, copy and remark of long-lived call object) ) To be. The scavenge GC period T is set as the target stop time− (program stop time due to mark, copy and remark of short-lived control signal object) / (program stop time due to mark, copy and remark of long-lived call object) This is because it is assumed that the life time of the short-lived control signal object is shorter than the scavenge GC period T.

  Note that the virtual machine tuning value calculation apparatus 10 may calculate a NEW area size (tuning parameter value) of each virtual machine in a system including a plurality of virtual machines (servers 20). For example, as shown in FIG. 4, a case is considered where the system is configured by servers 20 (servers 20a, 20b,..., 20n). In this case, the tuning parameter value calculation unit 12 of the virtual machine tuning value calculation apparatus 10 in FIG. 1 sets the target stop time of the entire system input via the input unit 11 for each server 20 (servers 20a, 20b,..., 20n). To distribute. That is, the target stop time of each server 20 is obtained from the target stop time of the entire system. The target stop time of each server 20 here is, for example, a value obtained by equally distributing the target stop time of the entire system to each server 20 (servers 20a, 20b,..., 20n) or the processing capacity of each server 20. A value distributed in accordance with this can be considered. The target stop time of each server 20 may be an arbitrary value set by the system developer as long as the total value becomes the target stop time of the entire system. Then, the tuning parameter value calculation unit 12 sets the target stop time of each of the servers 20 (servers 20a, 20b,..., 20n) and the application model information and traffic model information of each of the servers 20 (servers 20a, 20b,..., 20n). And the NEW region size (tuning parameter value) of each of the servers 20 (servers 20a, 20b,..., 20n) is calculated using the GC operation parameter information. In this way, when the system is composed of a plurality of virtual machines, the NEW area size (tuning parameter value) set for each virtual machine can be obtained.

  In the embodiment described above, the target stop time of the virtual machine (target value of program stop time during GC) is used when determining the NEW area size (optimum NEW area size) to be set in the virtual machine. Other conditions may be used.

  Generally, it is preferable that the program stop time at the time of GC of the virtual machine is as short as possible. Here, the program stop time is shortened when the NEW area size (especially, the size of the Eden area of the NEW area) is reduced, but it is not always necessary to simply reduce the NEW area (especially the size of the Eden area of the NEW area). . This is due to the following reason.

  In the above-described embodiment, it is assumed that the survival (holding) time of the short-lived control signal object is shorter than the scavenge GC cycle T. However, if the NEW area is too small, the NEW area is immediately filled with objects. As a result, the scavenge GC may be executed earlier than the average survival (holding) time of the short-lived control signal object. For this reason, the probability that the short-lived control signal object generated in the NEW area is copied to the OLD area by the scavenge GC is also increased. That is, in the scavenge GC, the copy time of the object from the NEW area to the OLD area increases, and when the number of objects copied to the OLD area increases in this way, the remark time in the CMS in the OLD area also increases. As a result, the program stop time increases. Further, if the NEW area size (particularly the size of the Eden area of the NEW area) is made too small as described above, the operating rate of the virtual machine is also significantly reduced. The operating rate here is the execution time of the application in the virtual machine / (the execution time of the application in the virtual machine + the scavenge GC stop time t1). That is, operation rate = scavenge GC cycle T / (scavenge GC cycle T + scavenge GC stop time t1).

  Therefore, in the calculation of the tuning value (for example, NEW area size) of the virtual machine, in addition to the program stop time during the GC of the virtual machine being equal to or less than the target stop time (D), the scavenge GC cycle T is a short-lived control signal. It is preferable to consider that it is longer than the object average hold time (B3) and that the operating rate of the virtual machine exceeds the target value.

  FIG. 5 is a diagram illustrating the operating rate of the virtual machine with respect to the NEW area size of the virtual machine. As shown in FIG. 5, the NEW area size requires that the operating rate of the virtual machine exceeds the lower limit (target value). Further, the short-lived control signal object does not encounter a plurality of GCs within the NEW region size range in which the operating rate of the virtual machine exceeds the lower limit (target value) (that is, scavenge GC cycle T> short-lived control). The range of the signal object average holding time (B3)) and the program stop time of the virtual machine less than the target stop time (D) is an appropriate NEW area size. That is, the range of S1 to S2 shown in FIG. 5 is an appropriate range of the NEW area.

  Therefore, the tuning parameter value calculation unit 12 of the virtual machine tuning value calculation apparatus 10 calculates an appropriate NEW region size using Equations (3) to (7) when calculating the NEW region size set in the virtual machine. The range may be calculated.

Target stop time (D)> scavenge GC stop time t1 + remark stop time t2 Formula (3)
Scavenge GC cycle T> short-lived control signal object average holding time (B3) Equation (4)
Scavenge GC cycle T / (scavenge GC cycle T + scavenge GC stop time t1)> target value of operating rate (5)
However,
Scavenge GC stop time t1 = (number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × short-lived control signal object average Holding time (B3) + scavenge GC cycle T × number of calls per unit time (A1) × number of long-lived call objects per call (B2)) × mark and copy per object in NEW area in scavenge GC Time (C2) ... Formula (6)
Remark stop time t2 = (number of calls per unit time (A1) × number of control signals per call (A2) × number of short-lived control signal objects per control signal (B1) × short-lived control signal object average hold Time (B3) + scavenge GC period × number of calls per unit time (A1) × number of long-lived call objects per call (B2)) × remark time per object copied to OLD area in CMS ( C3) ... Formula (7)

  It should be noted that the number of calls generated per unit time (A1) × the number of control signals per call (A2) × the number of short-lived control signal objects per control signal (B1) in Expression (6) and Expression (7) X Short-lived control signal object average holding time (B3)) indicates an average value of the number of short-lived control signal objects that survive in the NEW area during scavenge GC.

  Expression (6) and Expression (7) (scavenge GC period T × number of calls per unit time (A1) × number of long-lived call objects per call (B2)) occur during the scavenge GC period. Indicates the number of long-lived call objects. That is, it indicates the number of long-lived call objects that survive in the NEW area during scavenging GC.

  Note that, in a system composed of a plurality of virtual machines (servers 20), as described above, when determining an appropriate NEW area size range for each virtual machine, the target stop time of the entire system is set to the server 20 (servers 20a and 20b). , ..., 20n). Then, the tuning parameter value calculation unit 12 sets the target stop time of each of the servers 20 (servers 20a, 20b,..., 20n) and the application model information and traffic model information of each of the servers 20 (servers 20a, 20b,..., 20n). And the range of the appropriate NEW area size (tuning parameter value) for each of the servers 20 (servers 20a, 20b,..., 20n) may be calculated using the GC operation parameter information.

  In designing an application operating on a virtual machine, the size and number of each object affect the program stop time (the scavenge GC stop time t1 and the remark stop time t2 described above) due to GC. Therefore, setting the size and number of each object is important. That is, the settings of the number of short-lived control signal objects (B1), the size of short-lived control signal objects (B4), the number of long-lived call objects (B2), and the size of long-lived call objects (B5) are important. Here, if the target stop time (D) of the virtual machine, the target value of the operation rate, etc. are determined, the size and number of the short-lived control signal object and the long-lived call object can be determined by using the tuning value calculation algorithm described above. The range that can be taken is limited. For example, by using a tuning value calculation algorithm, the value of the number of short-lived control signal objects (B1) × the size of short-lived control signal objects (B4) and the long-lived, in which the program stop time of the virtual machine is less than the target stop time (D) The combination of the value of the number of call objects (B2) × the size of the long-lived call object (B5) is within the hatched range in FIG. Thus, the possible range of the value of the number of short-lived control signal objects (B2) × the size of the short-lived control signal object (B4) and the value of the number of long-lived call objects (B2) × the size of the long-lived call object (B5) can be known in advance. Thus, the system developer can easily design an application in which the program stop time is less than the target stop time (D).

DESCRIPTION OF SYMBOLS 10 Virtual machine tuning value calculation apparatus 11 Input part 12 Tuning parameter value calculation part 13 Output part 20 Server 111 Application model input part 112 Traffic model input part 113 GC operation parameter input part 114 Target stop time input part 131 Tuning parameter value output part

Claims (5)

A virtual machine tuning value calculation device for calculating a NEW area size of a memory set in a virtual machine used for a communication node,
The virtual machine, the virtual machine that the scavenge GC adopted as a method of GC (garbage collection) of NEW area on the memory, to adopt a CMS (Concurrent Mark & Sweep) as a method of GC of OLD region on the memory Yes,
The application used for the virtual machine manages a series of operations from when another terminal device accesses the virtual machine to the end of processing as a call related to a plurality of control signals, and as an object for the call. An application that generates one or more long-lived call objects and generates one or more short-lived control signal objects as objects for the control signal;
Traffic model information indicating the number of calls generated per unit time and the number of control signals per call in the application;
The number of short-lived control signal objects per control signal, the number of long-lived call objects per call, the average short-lived control signal object holding time, the size of the short-lived control signal object, and the size of the long-lived call object Application model information indicating
GC operation parameter information indicating the mark and copy time per object in the NEW area in the scavenge GC and the remark time per object copied to the OLD area in the CMS;
Receiving an input of a target stop time which is a target value of a program stop time at the time of GC in the virtual machine, and storing it in a storage unit;
Using the traffic model information, the application model information, the GC operation parameter information, and the target stop time of the storage unit,
A virtual machine tuning value calculation device comprising a tuning parameter value calculation unit for calculating the NEW region size according to the following equations (1) and (2).
NEW area size = scavenge GC cycle T × number of calls per unit time × (number of long-lived call objects per call × size of long-lived call object + number of control signals per call × short life per control signal Number of control signal objects × size of short-lived control signal object) (1)
However,
Scavenge GC cycle T = {target stop time− (number of calls per unit time × number of control signals per call × number of short-lived control signal objects per control signal × short-lived control signal object average holding time × scavenge GC Mark and copy time per object in the NEW area + number of calls generated per unit time x number of control signals per call x number of short-lived control signal objects per control signal x short-lived control signal object average hold time X Remark time per object copied to OLD area in CMS} / (Number of calls per unit time x Number of long-lived call objects per call x Mark per object in NEW area in scavenge GC And copy time + call per unit time Number of long-lived call objects per call × remark time per object copied to the OLD area in CMS) (2) When the tuning parameter value calculation unit calculates the NEW area size of each of the virtual machines in a system including a plurality of the virtual machines,
Using the traffic model information, the application model information, and the GC operation parameter information of each of the virtual machines, and the target stop time of each of the virtual machines that distributes the target stop time of the entire system,
The virtual machine tuning value calculation apparatus according to claim 1, wherein the NEW area size of each of the virtual machines is calculated by the expressions (1) and (2). The virtual machine tuning value calculation device is:
Further, it receives an input of the target value of the operating rate of the virtual machine, and stored in the storage unit,
When the tuning parameter value calculation unit calculates the NEW region size to be set in the virtual machine, the traffic model information, the application model information, the GC operation parameter information, the target stop time, and the storage unit are stored in the storage unit. Using the target value of the operating rate,
The following equation (3) Virtual Machine tuning value calculating device according to claim 1, wherein the benzalkonium to calculate the range of the NEW area size by to (7).
Target stop time> scavenge GC stop time t1 + remark stop time t2 Formula (3)
Scavenge GC period T> short-lived control signal object average holding time (4)
Scavenge GC cycle T / (scavenge GC cycle T + scavenge GC stop time t1)> target value of operating rate (5)
However,
Scavenge GC stop time t1 = (number of calls generated per unit time × number of control signals per call × number of short-lived control signal objects per control signal × short-lived control signal object average holding time + scavenge GC cycle T × unit Number of calls per time x number of long-lived call objects per call) x mark and copy time per object in NEW area in scavenge GC (6)
Remark stop time t2 = (number of calls per unit time × number of control signals per call × number of short-lived control signal objects per control signal × short-lived control signal object average hold time + scavenge GC cycle × per unit time Number of call occurrences x number of long-lived call objects per call) x remark time per object copied to OLD area in CMS (7) A virtual machine tuning value calculation method for calculating a NEW area size of a memory set in a virtual machine used for a communication node,
The virtual machine, the virtual machine that the scavenge GC adopted as a method of GC (garbage collection) of NEW area on the memory, to adopt a CMS (Concurrent Mark & Sweep) as a method of GC of OLD region on the memory Yes,
The application used for the virtual machine manages a series of operations from when another terminal device accesses the virtual machine to the end of processing as a call related to a plurality of control signals, and as an object for the call. An application that generates one or more long-lived call objects and generates one or more short-lived control signal objects as objects for the control signal;
A virtual machine tuning value calculation device for calculating the NEW region size is:
In the application, and call generation per unit of time, and traffic model information indicating the number of control signals per the call one,
The number of short-lived control signal objects per control signal, the number of long-lived call objects per call, the average short-lived control signal object holding time, the size of the short-lived control signal object, and the size of the long-lived call object Application model information indicating
GC operation parameter information indicating the mark and copy time per object in the NEW area in the scavenge GC and the remark time per object copied to the OLD area in the CMS;
Receiving an input of a target stop time that is a target value of a program stop time during GC in the virtual machine and storing the input in a storage unit;
Using the traffic model information, the application model information, the GC operation parameter information, and the target stop time of the storage unit,
And a step of calculating the NEW region size according to the following formulas (1) and (2):
NEW area size = scavenge GC cycle T × (number of calls generated per unit time × number of long-lived call objects per call × size of long-lived call object + number of calls generated per unit time × number of control signals per call X number of short-lived control signal objects per control signal x short-lived control signal object average holding time x size of short-lived control signal object) (1)
However,
Scavenge GC cycle T = {target stop time− (number of calls per unit time × number of control signals per call × number of short-lived control signal objects per control signal × short-lived control signal object average holding time × scavenge GC Mark and copy time per object in the NEW area + number of calls generated per unit time x number of control signals per call x number of short-lived control signal objects per control signal x short-lived control signal object average hold time X Remark time per object copied to OLD area in CMS} / (Number of calls per unit time x Number of long-lived call objects per call x Mark per object in NEW area in scavenge GC And copy time + call per unit time Number of long-lived call objects per call × remark time per object copied to the OLD area in CMS) (2) The virtual machine tuning value calculation device is:
Further, it receives an input of the target value of the operating rate of the virtual machine, and executes the step of storing in the storage unit,
When executing the step of calculating the NEW region size, using the traffic model information, the application model information, the GC operation parameter information, the target stop time and the target value of the operation rate in the storage unit,
Virtual Machine tuning value calculation method according to claim 4, characterized in that to perform the steps of calculating the range of the NEW region size according to the following equation (3) to (7).
Target stop time> scavenge GC stop time t1 + remark stop time t2 Formula (3)
Scavenge GC period T> short-lived control signal object average holding time (4)
Scavenge GC cycle T / (scavenge GC cycle T + scavenge GC stop time t1)> target value of operating rate (5)
However,
Scavenge GC stop time t1 = (number of calls generated per unit time × number of control signals per call × number of short-lived control signal objects per control signal × short-lived control signal object average holding time + scavenge GC cycle T × unit Number of calls per time x number of long-lived call objects per call) x mark and copy time per object in NEW area in scavenge GC (6)
Remark stop time t2 = (number of calls per unit time × number of control signals per call × number of short-lived control signal objects per control signal × short-lived control signal object average hold time + scavenge GC cycle × per unit time Number of call occurrences x number of long-lived call objects per call) x remark time per object copied to OLD area in CMS (7)

Images