SYSTEM AND METHOD FOR IMPROVING MEMORY USAGE IN VIRTUAL MACHINES

[0001] This application is a Continuation application of U.S. patent application Ser. No. 13/738,814, filed on Jan. 10, 2013.

[0002] 1. Field of the Invention

[0003] The present invention relates generally to memory optimization and improving the efficiency of real memory use by applications and computing platforms such as physical and virtual machines. More specifically, in a system having a plurality of virtual machines, information is periodically collected on memory and CPU usage of each virtual machine and a memory optimizer uses this information to re-allocate memory among the virtual machines, as based on at least one memory optimization mechanism which can reduce memory usage of a virtual machine at a cost of increasing CPU usage.

[0004] 2. Background of the Invention

[0005] It is often important to run multiple virtual machines (VMs) concurrently, particularly in cloud computing environments, where multiple applications can be run concurrently. However, poor memory usage can make a system almost unusable, as, for example, when too much paging between memory and disk can slow the system down.

[0006] In view of the foregoing and other exemplary problems, drawbacks, and disadvantages of conventional methods and systems, an exemplary feature of the present invention is to provide a method and structure to dynamically allocate memory between concurrently-running virtual machines.

[0007] In a first exemplary aspect of the present invention, described herein is an apparatus, including: at least one processor upon which can be executed a virtual memory optimizer for optimizing a memory usage among a plurality of concurrently-running virtual machines; and a memory that stores a set of computer readable instructions for implementing and executing the virtual memory optimizer, the memory optimizer performing a monitoring of a usage of memory by each virtual machine of the plurality of concurrently-running virtual machines and applying at least one memory optimization mechanism that reallocates memory among the concurrently-running virtual machines based on reducing a memory usage of a virtual machine as a tradeoff of increasing a central processing unit (CPU) usage to achieve the reduced memory usage.

[0008] In a second exemplary aspect of the present invention, also described herein is a system comprising a plurality of concurrently-running virtual machines (VMs), the system comprising at least one computer including: at least one processor upon which can be executed a virtual memory optimizer for optimizing a memory usage among the plurality of concurrently-running VMs; and a memory that stores a set of computer readable instructions for implementing and executing the virtual memory optimizer, the virtual memory optimizer providing at least one memory optimization mechanism which can reduce a memory usage of a virtual machine at a cost of increasing a central processing unit (CPU) usage.

[0016] FIG. 1 depicts an exemplary system having features of the present invention. A memory optimizer 102 optimizes memory use across multiple virtual machines 101. The memory optimizer could be implemented in software and can execute on one or multiple processor nodes. The virtual machines 101 may also execute on one or more processor nodes.

[0017] Virtual machines may be implemented in multiple ways including but not limited to IBM's Dynamic Logical Partitioning (DLPAR) (described in various publications, including an internal IBM web page identified in the cited reference section on the front page of the issued patent version of this disclosure. This publication explains how DLPAR provides the ability to logically attach and detach a managed system's resources to and from a logical partition's operating system without rebooting. The contents of this publication are incorporated herein by reference), VMware virtualization platforms such as ESX, Xen, KVM, Java virtual machines, etc. Information on commonly used virtualization platforms is available in a number of publications, including a generic description in a wikipedia article entitled “Comparison_of_platform_virtual_— machines”. The contents of this publication are incorporated herein by reference.

[0018] This publication explains how platform virtual machines are software packages that emulate a whole physical computer machine, often providing multiple virtual machines on one physical platform. This publication also provides a comparison of basic information about a relatively large number of platform virtual machine (VM) packages.

[0019] The memory optimizer 102 of the present invention could run on one or more nodes on which the virtual machines 101 execute. Alternatively, the memory optimizer 102 could run on one or more nodes which are distinct from the one or more nodes on which the virtual machines 101 execute. Alternatively, the memory optimizer 102 could run one or more nodes on which the virtual machines 101 execute as well as one or more nodes which are distinct from the one or more nodes on which the virtual machines 101 execute.

[0020] The piece of software, firmware, or hardware that creates and runs virtual machines is often referred to as a hypervisor. All or a fraction of the memory optimizer 102 could be part of a hypervisor. However, this is not necessary. The memory optimizer 102 can also execute independently from a hypervisor.

[0021] In the context of explaining the present invention, there is a pool of memory that can be allocated to multiple virtual machines. In other words, some memory m1 could be allocated to either virtual machine vm1 or virtual machine vm2. If vm1 has a lot of free memory while vm2 has little free memory, it is preferable to allocate m1 to vm2. By contrast, if vm2 has a lot of free memory while vm1 has little free memory, it is preferable to allocate m1 to vm1.

[0022] The memory optimizer 102 determines how memory should be allocated to different virtual machines. If vm1 has a lot of free memory while vm2 is running low on free memory, memory optimizer 102 might allocate some of vm1's free memory to vm2.

[0023] Once a virtual machine 101 starts running low on memory, its performance can degrade considerably. In some cases, insufficient memory can cause applications to not execute properly at all. It is therefore desirable to maintain proper amounts of memory for all virtual machines 101. If sufficient memory exists throughout the system, the memory optimizer 102 can allocate memory across the virtual machines 101 to provide enough memory for each virtual machine 101. If there is not enough memory in the system to prevent a virtual machine 101 from running low on memory, other actions need to be taken.

[0024] There are various special techniques that can be deployed to reduce the memory consumed by a virtual machine. One such technique is memory compression. Another technique is memory deduplication, in which duplicate memory pages are identified. Memory deduplication stores a single copy of duplicate memory pages. A third technique is delta encoding. For two pages p1 and p2 which are similar but not identical, it is not necessary to store entire versions of both p1 and p2. Instead, p1 could be stored along with a delta which encodes just the differences between p1 and p2. The delta would typically be considerably smaller than p2.

[0025] The use of these and other techniques for virtual machines is described in D. Gupta et al, “Difference Engine: Harnessing Memory Redundancy in Virtual Machines”, Proceedings of OSDI 2008, the contents of which is incorporated herein by reference. There are other techniques that can be applied to reduce memory usage as well. For example, a virtual machine 101 might be a Java virtual machine (abbreviated JVM) running a Java workload. An overview of JVMs is described in various publications, including a Wikipedia article “Java_virtual_— machine”, the content of which is incorporated herein by reference. Java allocates memory from an area known as the heap. In order to free memory from the heap, garbage collection is needed.

[0026] The concepts of “heap” and “garbage collection” are further explained in various publications, such as the Wikipedia article entitled “programming_language”, the Wikipedia article entitled “memory management”, and the Wikipedia article entitled “garbage collection”, the contents of these articles being hereby incorporated herein by reference. In summary, memory management is the process of managing computer memory, including mechanisms to dynamically allocate portions of memory to programs upon request and freeing it for reuse when no longer needed. Memory requests are satisfied by allocating portions from a large pool of memory referred to as “the heap”, and, at any given time, some parts of the heap are in use while some are “free” (i.e., unused) and thus available for future allocations.

[0027] Garbage collection is a form of automatic memory management that attempts to reclaim “garbage”, as referring to memory occupied by objects that are no longer in use by a program. Garbage collection is often portrayed as the opposite of manual memory management, a mechanism in which the programmer specifies which objects to deallocate and return to the memory system. Many computer languages, such as Java, C#, and several scripting languages, require garbage collection either as part of the language specification or effectively for practical implementation, while other languages, such as C, C++, were designed for use with manual memory management but have garbage collected implementations available. Still others, such as Ada, Modula-3, and C++/CLI allow both garbage collection and manual memory management to co-exist in the same application by using separate heaps, and others, such as D, are garbage collected but allow the user to manually delete objects and disable garbage collection when speed is required.

[0028] The entity that performs garbage collection is known as the garbage collector. Garbage collectors are typically implemented in software, but they can also be implemented using both software and hardware. For example, a system might have special hardware support to aid garbage collection.

[0029] Returning now to an exemplary embodiment of the present invention using JVM, if a smaller maximum heap size is used, the JVM consumes less memory. However, the garbage collector needs to run more frequently, which uses up additional CPU cycles. Thus, there is once again a memory/CPU trade-off. Larger maximum heap sizes use up more memory for the heap but save CPU cycles because less frequent garbage collections are needed. Smaller maximum heap sizes use up less memory but use more CPU cycles because more frequent garbage collections are needed.

[0030] Note that this technique of modifying Java heap space is applicable to other languages with automatic memory management and garbage collection, such as Lisp, Smalltalk, C#, many scripting languages, etc. The present invention is applicable to languages which use garbage collection in general. The term “heap” refers to the memory area managed using garbage collection and is not specific to the Java programming language. However, for ease of exposition, we exemplarily discuss concepts of the present invention in terms of Java and Java virtual machines. One skilled in the art could easily apply this invention to other languages with garbage collection.

[0031] Other techniques besides those mentioned above can be used for optimizing memory usage within the spirit and scope of this invention.

[0032] As recognized by the present inventors, a key problem with these techniques is that they consume CPU overhead. Thus, while they improve memory usage, they hurt CPU performance. The present inventors have recognized that what is needed is a selective way to apply these techniques. That is what the present invention provides,

[0033] Thus, in the present invention, memory optimizer 102 shown in FIG. 1 takes a global view of all virtual machines 101 that potentially can be allocated memory under its control, as follows. When free memory throughout the system is low and a particular virtual machine 101 needs more memory, the memory optimizer identifies one or more virtual machines 101 which has spare CPU capacity to apply an optimization to free up memory. That way, a virtual machine 101 running on hardware with little spare CPU capacity will not have to devote additional CPU cycles in order to free up memory.

[0034] FIG. 2 depicts an exemplary method for managing memory across multiple virtual machines 101. In Step 201, the memory optimizer 102 is continuously monitoring the system. It collects information such as memory usage of virtual machines, CPU usage of virtual machines, etc. For such information collection, memory optimizer 102 can use standard tools (e.g. vmstat) in operating systems such as Linux, other variants of Unix, Windows, etc. to collect this information. Other methods and tools can also be used to monitor memory and CPU usage within the spirit and scope of this invention.

[0035] In step 202, memory optimizer 102 determines that a particular virtual machine VMx (e.g., VM1) needs more memory. This can be done in several ways.

[0036] For example, memory optimizer 102 might determine that the amount of free memory available to VM1 is decreasing and getting close to 0 (memory which is available to a virtual machine 101 but is not being used is “free memory”); more specifically, the amount of free memory may fall below a threshold. Alternatively, the memory optimizer 102 might determine that VM1 has no free memory and is paging. Alternatively, the memory optimizer 102 might have predictions of future memory needs for VM1 based on empirical data of VM1's memory usage in the past. Even though VM1 currently has some free memory, the memory optimizer 102 might predict that VM1 is likely to run out of memory in the near future unless it is given more memory. Other methods for determining that VM1 needs more memory are possible within the spirit and scope of the invention.

[0037] The memory optimizer 102 has a global view of how much memory each virtual machine 101 has available to it and how much memory a virtual machine 101 is using. If, in step 203, the memory optimizer 102 determines that sufficient free memory exists from other virtual machines 101 to satisfy the memory needs of VM1, the memory optimizer 102, in step 204, allocates additional memory to VM1 from one or more other virtual machines 101 with free memory.

[0038] If the memory optimizer 102 determines in step 203 that sufficient free memory from other virtual machines 101 does not exist to fully satisfy the memory needs of VM1, the memory optimizer in step 205 tries to identify one or more virtual machines 101 which can free up memory to give to VM1 by applying an optimization. Such optimizations include but are not limited to the aforementioned memory compression, memory deduplication, delta encoding, and reducing heap space for Java applications and/or applications in other programming languages with automatic memory management and garbage collection.

[0039] The memory optimizer 102 uses knowledge of specific optimizations to estimate both the CPU overhead of applying an optimization and the amount of memory which would be freed by applying the optimization. For example, information on memory used by specific applications can be analyzed to determine both the amount of memory saved and the CPU overhead incurred for applying optimizations such as memory compression, memory deduplication, and delta encoding.

[0040] A specific example of a possible mechanism the memory optimizer 102 could use to estimate both the CPU overhead of applying an optimization and the amount of memory which could be freed by applying the optimization is IBM's amepat tool (Active Memory™ Expansion Planning and Advisory Tool), described in various publications including internal IBM publication located at the URL address identified in the references section. The content of this publication, incorporated herein by reference, describes amepat. Amepat provides information on memory saved and CPU overhead incurred by memory compaction.

[0041] Another example of how the overhead of applying an optimization and the amount of memory which could be freed by applying the optimization could be estimated is the following. For programming languages using garbage collection, such as Java, empirical data can be collected on overhead incurred by garbage collection as a function of heap size. If memory usage of individual applications is profiled, these estimates of garbage collection overheads can be more accurate. From this data, the memory optimizer 102 will be able to estimate the effect of heap size on garbage collection overhead. If a virtual machine has sufficient excess CPU capacity, the memory optimizer 102 can reduce the Java heap size appropriately to free up memory.

[0042] Based on estimates of CPU overhead incurred for an optimization and the memory expected to be freed, the memory optimizer 102 determines which optimization (s) should be applied to which virtual machines 101 to free memory. The CPU load on the virtual machines 101 is an important part of the decision. If a virtual machine 101 has high CPU utilization or is predicted to have high CPU utilization in the near future, that virtual machine 101 is not a good candidate to which to apply a CPU-intensive memory optimization. On the other hand, if a virtual machine 101 is consuming few CPU cycles and is not predicted to have a significant increase in CPU consumption in the near future, that virtual machine might be a better candidate for applying the memory optimization, provided the virtual machine is predicted to release a significant amount of memory as a result of applying the optimization.

[0043] Memory usage by a virtual machine 101 is also used to determine whether it is a good candidate for giving up memory. If a virtual machine 101 is not using up much memory, then it is not a good candidate for giving up memory by applying an optimization since it has little memory to give. If, on the other hand, a virtual machine 101 is using up a lot of memory which could be freed by applying an optimization, then it is a good candidate for giving up memory by applying an optimization.

[0044] There are several different criteria which can be applied to determine which optimizations should be applied to which virtual machine, including, but not limited, to the following:

• • [0045] If a virtual machine's predicted CPU utilization after applying the optimization is below a threshold value (e.g. 0.5), apply the optimization.
  • [0046] If a virtual machine's predicted CPU utilization after applying the optimization is below a threshold value and the amount of memory predicted to be freed by the optimization exceeds a threshold (e.g. 100 Megabytes), apply the optimization.
  • [0047] If a virtual machine's predicted CPU utilization after applying the optimization is below a threshold value and the amount of memory predicted to be freed by the optimization exceeds a threshold proportion of the memory needed by VM1, (e.g. 0.05), apply the optimization.
  • [0048] Each virtual machine has a target CPU utilization. Give priority to applying optimizations to virtual machines which have the highest value of target CPU utilization minus actual CPU utilization.
  • [0049] Each virtual machine has a target memory usage. Give priority to applying optimizations to virtual machines which have the lowest values of target memory utilization minus actual memory utilization.

[0050] It should be mentioned that each of the thresholds mentioned above could be different for different virtual machines 101. Alternatively, one or more thresholds could be the same for one or more virtual machines.

[0051] The memory optimizer 102 may apply one or more of the criteria above in determining which optimizations to apply to which virtual machines 101. For example, the memory optimizer 102 could give priority to applying optimizations to virtual machines which have the highest value of target CPU utilization minus actual CPU utilization while also giving priority to applying optimizations to virtual machines which have the lowest values of target memory utilization minus actual memory utilization.

[0052] It may be possible to apply multiple different optimizations to free up memory. In this case, memory optimizer 102 makes intelligent choices of which optimization (s) to apply, and to what degree. Thus, in an exemplary embodiment of the present invention, memory optimizer 102 can make the choices of which optimization (s) to apply (or increase use of) using the following guidelines:

• • [0053] Give priority to optimizations which consume relatively few CPU cycles.
  • [0054] Give priority to optimizations which result in relatively more memory savings.
  • [0055] Consider both the memory saved by an optimization and its CPU overhead in determining which optimization (s) to apply (or increase use of). An optimization is a good candidate if it is both relatively inexpensive in CPU cost and saves a relatively high amount of memory compared with other optimizations.

[0056] Memory optimizer 102 can also use other guidelines to determine which optimization (s) to apply or to increase use of.

[0057] In step 206, one or more optimizations are applied (and/or the application of one or more optimizations currently being used is increased). The memory freed by the one or more optimizations is given to virtual machine vm1.

[0058] The memory optimizer 102 also has the ability to reduce (or eliminate use of) the amount of an optimization applied to a virtual machine 101 if the virtual machine 101 is consuming too many CPU cycles. This is illustrated in FIG. 3.

[0059] In step 201 of FIG. 3, the memory optimizer monitors the system. This is the same step as step 201 in FIG. 2. In step 301, memory optimizer 102 detects a virtual machine VMx (e.g., VM2) which is consuming too many CPU cycles. For example, VM2 might have a CPU utilization exceeding a predetermined threshold, such as 90%. CPU utilization statistics may be determined from one or more measurements. In the case of multiple CPU utilization measurements, an aggregate figure, such as an average CPU utilization, may be determined. In some cases, step 301 might determine that even though VM2 is not currently consuming too many CPU cycles, it is likely to do so in the near future based on an analysis of past behavior, and action must be taken now to prevent CPU overload from occurring. The memory optimizer 102 then determines what optimizations, if any, are being applied by VM2. If it detects optimizations being applied to VM2, it reduces (or eliminates) the level of one or more optimizations to reduce CPU usage in step 302. As mentioned above, optimizations might include, but are not limited to, memory compression, memory deduplication, delta encoding, reducing heap space for Java applications and/or applications in other programming languages with automatic memory management and garbage collection. For example, the optimizer could increase the size of a Java heap in step 302 to reduce CPU cycles consumed by garbage collection.

[0060] There are multiple methods by which memory optimizer 102 could reduce (or eliminate) the amount of one or more optimizations applied to VM2. These include but are not limited to the following:

• • [0061] VM2 has a threshold CPU utilization (e.g. 0.5) which is being exceeded. Reduce (or eliminate) optimizations to bring the CPU utilization for VM2 within the threshold CPU utilization.
  • [0062] Multiple optimizations are being applied to VM2. Give priority to reducing (or eliminating) optimizations which consume significant CPU cycles.
  • [0063] Multiple optimizations are being applied to VM2. Give priority to reducing (or eliminating) optimizations which result in less memory saved.
  • [0064] Multiple optimizations are being applied to VM2. Consider both the memory saved by an optimization and its CPU overhead in determining which optimization (s) to reduce. An optimization is a good candidate to reduce (or eliminate) if it is both relatively expensive in CPU cost and saves relatively little compared with other optimizations.

[0065] Exemplary Hardware Implementation

[0066] From the exemplary embodiments described above, it is clear that the present invention is directed to controlling memory allocation for virtual machines and involves control concepts that would typically involve software. However, as is well known in the art, software implementation inherently involves underlying hardware. FIG. 4 exemplarily demonstrates a computer system 400 upon which a plurality of virtual machines operate using a hypervisor 401 that allocates and controls machine assets for a plurality of LPARs (logical partitions) upon which are respectively running VMs 402. This system 400 might involve a single computer or could be implemented across a network of interconnected computers. As pointed out about, exemplary embodiments of the present invention involve DLPAR, in which computer resources can be dynamically reallocated without rebooting.

[0067] FIG. 5 illustrates a typical hardware configuration of an information handling/computer system in accordance with the invention and which preferably has at least one processor or central processing unit (CPU) 511. In the context of the present invention, this computer would typically be a server on a network.

[0068] The CPUs 511 are interconnected via a system bus 512 to a random access memory (RAM) 514, read-only memory (ROM) 516, input/output (I/O) adapter 518 (for connecting peripheral devices such as disk units 521 and tape drives 540 to the bus 512), user interface adapter 522 (for connecting a keyboard 524, mouse 526, speaker 528, microphone 532, and/or other user interface device to the bus 512), a communication adapter 534 for connecting an information handling system to a data processing network, the Internet, an Intranet, a personal area network (PAN), etc., and a display adapter 536 for connecting the bus 512 to a display device 538 and/or printer 539 (e.g., a digital printer or the like).

[0069] In addition to the hardware/software environment described above, a different aspect of the invention includes a computer-implemented method for performing the above method. As an example, this method may be implemented in the particular environment discussed above.

[0070] Such a method may be implemented, for example, by operating a computer, as embodied by a digital data processing apparatus, to execute a sequence of machine-readable instructions. These instructions may reside in various types of non-transitory signal-bearing storage media.

[0071] Thus, this aspect of the present invention is directed to a programmed product, comprising non-transitory signal-bearing storage media tangibly embodying a program of machine-readable instructions executable by a digital data processor incorporating the CPU 511 and hardware above, to perform the method of the invention.

[0072] This signal-bearing storage media may include, for example, a RAM device 514 contained within the CPU 511, as represented by the fast-access storage, for example, and used for programs being currently executed, or a ROM device 516 storing program instructions not currently being executed. Alternatively, the instructions may be contained in another signal-bearing storage media, such as a magnetic data storage diskette 600 (FIG. 6), directly or indirectly accessible by the CPU 511.

[0073] Whether contained in the diskette 600, the computer/CPU 511, or elsewhere, the instructions may be stored on a variety of machine-readable data storage media, such as DASD storage (e.g., a conventional “hard drive” or a RAID array), magnetic tape, electronic read-only memory (e.g., ROM, EPROM, or EEPROM), an optical storage device (e.g. CD-ROM, WORM, DVD, digital optical tape, etc.), paper “punch” cards, or other suitable signal-bearing storage media including memory devices in transmission hardware, communication links, and wireless, and including different formats such as digital and analog. In an illustrative embodiment of the invention, the machine-readable instructions may comprise software object code.

[0074] As is readily apparent from the above description, the present invention discusses a new method for dynamically controlling memory allocation for a plurality of concurrently-operating VM's, as based on applying one or more memory optimization mechanisms for reducing memory usage of a VM taking into account a cost of increasing CPU usage. The method also permits CPU usage to be controlled, as related to memory usage.

[0075] Although the present invention has been described in various exemplary embodiments, it should be apparent that variations of this exemplary embodiment are possible and considered as included in the present invention.

[0076] Therefore, it is noted that, Applicants' intent is to encompass equivalents of all claim elements, even if amended later during prosecution.