Tuning adaptive computations for the performance improvement of applications in JEE server
- First Online:
- Cite this article as:
- Zhang, Y., Huang, G., Liu, X. et al. J Internet Serv Appl (2012) 3: 143. doi:10.1007/s13174-012-0060-4
- 426 Downloads
With the increasing use of autonomic computing technologies, a Java Enterprise Edition (JEE) application server is implemented with more and more adaptive computations for self-managing the Middleware as well as its hosted applications. However, these adaptive computations consume resources such as CPU and memory, and can interfere with the normal business processing of applications at runtime due to resource competition, especially when the whole system is under heavy load. Tuning these adaptive computations from the perspective of resource management becomes necessary. In this article, we propose a tuning model for adaptive computations. Based on the model, tuning is carried out dynamically by upgrading or degrading the autonomic level of an adaptive computation so as to control its resource consumption. We implement the RSpring tuner and use it to optimize autonomic JEE servers such as PkuAS and JOnAS. RSpring is evaluated on ECperf and RUBiS benchmark applications. The results show that it can effectively improve the application performance by 13.6 % in PkuAS and 19.2 % in JOnAS with the same amount of resources.
KeywordsAutonomic JEE server Adaptive computation Resource management
Being a runtime infrastructure between operating systems and JEE applications, today’s JEE servers increase their autonomic management capabilities by implementing more and more adaptive computations [1, 2, 3, 4]. According to the autonomic computing white paper , the term “adaptive computation” (AC) is used to denote any computing activity that implements either the complete autonomic control loop, or just one or more functions in the loop, i.e., monitor, analyze, plan, or execute. A JEE server could have tens or hundreds of adaptive computations because a lot of software/hardware objects need to be managed automatically by the server . For example, PkuAS  has been equipped with tens adaptive computations, such as automatically adjusting its thread-pool to guarantee the predefined tradeoff between the throughputs and response times , recovering from the correlated faults occurred in its internal services , changing the execution flow of a hosted application for trading off the performance and security .
The adaptive computations increase the performance, dependability, and flexibility of JEE servers. However, their runtime resource cost cannot be ignored, especially as the number of these computations keeps growing. An adaptive computation can collect the states of the managed objects (i.e., software components or hardware facilities); analyze the collected data to decide whether a system change is necessary; plan when and how to change; and finally perform the change on the fly. Obviously, each of the above four activities and their combinations consume resources, e.g., CPU and memory. A JEE application can run only when being deployed in a JEE server. Thus, it brings challenges to resource allocation for achieving the overall system performance goals because of the always existed resource competition between the adaptive computations of the JEE server and the business functions of the deployed applications. When the whole system is under stress, such a competition will cause negative effects on the applications and adaptive computations, and finally decrease the performance of the whole system.
There are generally two ways to mitigate the above problem. One is to increase the resources . For instance, redeploying the JEE server as well as its applications to the newly bought or rented machines. Such a solution is usually accompanied by increased operating costs. The newly allocated resources may become insufficient or excessive after a while under the quickly fluctuated system workloads. Therefore, there exists another solution to the problem by changing the quota of the current available resources between business functions and adaptive computations . This solution tries to exploit and reallocate the resources so as to make a quick response to the changing workloads. The above two solutions are complementary. However, when targeting a resource limited environment, we focus more on the latter in this article due to the desire to make full and flexible use of the resources that a JEE server currently has.
The runtime cost of adaptive computations leads to an important resource management issue, i.e., how to allocate the current available resources among adaptive computations and business functions, as well as among multiple adaptive computations. Compared with business functions, adaptive computations have their own specific features, such as the monitor-analyze-plan-execute autonomic loop, and the objects to be managed, e.g., thread-pool. Such features imply the rationales of resource management for adaptive computations. For instance, an adaptive computation can be “natively” upgraded/degraded rather than be roughly started/stopped, e.g., be degraded to run only the monitor function or be upgraded to run through all the four functions in the current execution of the autonomic loop. In that sense, adaptive computations can be controlled to execute flexibly to spare resources for processing more business requests. On the other hand, the running of some adaptive computations should be guaranteed because they are critical to the system. For example, the computations for self-healing and self-protecting can prevent the damages of abnormal and malicious conditions. Such computations can bring more gains or the so called cost-effectiveness to the system than the noncritical ones (Sect. 2.2). We cannot reclaim the resources consumed by these critical adaptive computations even if the resources are not enough for business functions.
In general, the core of a tuning solution to the above resource management issue includes two parts, i.e., model and implementation. The model answers the question of how to allocate resources based on the abstraction of computing entities; the implementation answers when and to what extent the resources are allocated to and consumed by each computing entity. Take the classic process scheduling as an example, a process is scheduled to consume a certain amount of resources, e.g., CPU. The resource management model contains the following elements: the process life-cycle abstraction (e.g., start/stop), the process priority, etc. Based on the model, the implementations decide how exactly each specific process will be executed with a given amount of resources, e.g., make a process stop/suspend to return the possessed CPU cycles; or make a process run more often than the others so as to consume more CPU cycles. For achieving the best-of-the-breed effectiveness and efficiency, the model and implementations always leverage the features of the computing entities. Therefore, these two parts usually make the resource management solutions specific to the target system. In an autonomic JEE server, we observe that meeting the resource needs of business functions is in most cases more important than that of the noncritical adaptive computations, and we accept the notion that business requirements should be satisfied first when fierce resource competition occurs. Therefore, to enable a flexible tradeoff between business functions and adaptive computations when resources are limited and competed, what can be done is: if necessary, executing adaptive computations at a right time to a proper extent. We have demonstrated its possibility in the position paper . In this article, we focus on the tuning model specific to adaptive computations, implement the RSpring tuner for autonomic JEE servers, and perform a thorough evaluation on RSpring using industry standard benchmarks.
A tuning model for controlling the resource allocation between business functions and adaptive computations by upgrading and degrading the autonomic levels of adaptive computations dynamically. The features and gains of adaptive computations have been taken into considered in the model.
The RSpring tuner for autonomic JEE servers such as PkuAS1 and JOnAS.2 A set of algorithms has been built in the tuner for determining how exactly the autonomic level of each adaptive computation is tuned.
We evaluate RSpring across a range of experiments with ECperf3 and RUBiS4 JEE benchmarks. The results show that RSpring is effective and efficient to support the performance of the whole JEE system, and in particular, our considerations of the features and gains of adaptive computations in model- and algorithm-design are valid.
The rest of this article is organized as follows: Section 2 gives an overview of adaptive computations and a real case illustrating the resource competition problem. Section 3 describes the tuning model for adaptive computations. Section 4 presents the RSpring tuner and its built-in tuning algorithms. Section 5 reports the evaluation on RSpring. Section 6 compares our work with related efforts. Section 7 ends this article with conclusion and future work.
2 Background and a motivating example
2.1 Adaptive computation
(AC) is the abstract representation of the autonomic control loop described above. While a so called mature AC  should include all the four functions in the loop, other ACs can include only parts of the four functions. For instance, a JEE server such as PkuAS provides several JMX MBean services  that expose the internal states of the server and the hosted applications in its management console. Such a service can be considered as an AC with only the monitor function. Another case is LTA (the “Log/Trace Analyzer,” a component in the IBM autonomic computing toolkit: Tivoli), which can be considered as an AC with monitor and analyze functions.
The autonomic level  of an AC refers to the maximum allowed completion level in the current iteration of its control loop. For instance, if LTA is allowed to run only the monitor function in an iteration of its control loop, and is allowed to run through all the two functions (i.e., monitor and analyze) in the next iteration, we say that the autonomic level of LTA in the former case is lower than that in the latter case. The level-changing from the former to the latter is called upgrade, otherwise, degrade. Upgrading the autonomic level of an AC is usually accompanied by increased resource consumption because more functions can be run, while degrading the level is usually accompanied by decreased resource consumption.
2.2 A motivating example
PkuAS  is an open source JEE server. It has several ACs in the form of JMX MBean services such as Response Time Monitor (RTM), Log Processor (LP), and Thread Pool Adjuster (TPA) . RTM monitors the response time of each client request and presents the results in the management console. LP handles the logs. It continuously filters and aggregates the concerned data in each log file, and stores them into a database for further processing. For instance, if a serious system error appears in a log repeatedly in a very short time, LP will mark this error, put it into the database, and then send an alarm to administrators. TPA dynamically adjusts the thread-pool size of PkuAS in accordance with response time and throughput .
ECperf  is a JEE benchmark application. Its EJB components perform the typical e-business transactions (e.g., products-browsing and ordering). Its test driver simulates client requests to ECperf’s EJBs deployed in a JEE server. The driver has an adjustable integer parameter called txRate, which determines the requests generation rate. 1 txRate can be considered as 1,000 transactions per minute approximately. Thus, the workload increases with the increase of txRate.
When txRate is 1 and 2, the available resources are relatively abundant with regard to the resource needs of ACs and business functions at the given workloads. Thus, we get the same “response time of 90 % NewOrder” results in the “No AC Runs” and “ACs Run” tests. However, when txRate is above 3, the resources become scarce at the given workloads. In such cases, the ACs indeed compete with business functions for the limited resources, which will surely interfere with the processing of the latter. For instance, when txRate is 6, the tested result in “ACs Run” is 141 % of that in “No AC Runs” test. ACs can incur performance penalties, but being a full-fledged autonomic JEE server, PkuAS needs to run together with them. For instance, LP manages logs and RTM records response time for administration. As to TPA, what should be noted is that it manages the thread-pool of PkuAS, which can improve the system performance if the management effects exceed the runtime costs. That is why we see in Fig. 2 that when txRate is above 6, the performance of PkuAS running with ACs is much better than that of the “pure” PkuAS. The “ACs Run” curve also illustrates the main use of ACs: self-handling unexpected extreme cases such as a sudden surge of client requests or runtime exceptions.
The test of “ACs Run (stop RTM at 6)” in Fig. 2 proves the worth of degrading or stopping the noncritical ACs to support system performance when necessary, that is, degrade or stop the ACs that contribute little to system performance improvement, and meanwhile have little impact on the correct execution of the system if being degraded or stopped. The curve shows that the system performance in this case is even better than that of the “ACs Run” test when txRate≥6. For instance, when txRate is 10, the response time of “90 % NewOrder” is 10.8 % lower, which is a great improvement when under such a heavy workload. However, the manual tuning way may be ineffective to achieve timely response to varying workloads. Therefore, an automatic tuner is necessary.
In addition, different ACs should be treated differently when being tuned. Each AC has its own specific features and usually produces different gains with regard to system performance. For example, if PkuAS does not run with TPA, its performance could be poorer most of the time than running with TPA together. TPA can produce greater gains and is more cost-effective than the other two ACs with regard to system performance. Thus, it should better be run or its autonomic level should be upgraded first if resources are available to do so. On the other hand, when executed under an extremely heavy load, TPA may be less effective than in usual cases. Because only adjusting the thread-pool cannot mitigate the increasingly fierce resource competition between business functions and ACs, and cannot offset the deterioration of system performance caused by the heavy workloads. In such a case, a better strategy can be to degrade or stop TPA for devoting its resource-holdings to business functions. Thus AC tuning should have a global view, and should consider the specific features and gain of each AC.
3 Tuning model
We propose an autonomic control loop-based model to tune the resource allocation between business functions and ACs by upgrading and degrading the autonomic levels of ACs dynamically. The model considers the features and gain of an AC. Based on the model, the RSpring tuner is implemented for tuning the ACs of an autonomic JEE server when resources are limited and competed.
3.1 The features of an adaptive computation
Tuning an AC
Time driven; Tuner driven
Time/monitor driven; Tuner driven
Analyzer driven; Tuner driven
Planner/analyzer driven; Tuner driven
Start/Stop, or tune the execution cost
The valid states of an AC
All the four autonomic functions in the control loop are stopped
Only the monitor function is started
S1 is true and the analyze function is started
S2 is true and if a change is necessary, the execute function is started
As mentioned in Sect. 2.1, not all the four functions must exist in an AC, thus an AC may only have parts of the states listed above. In addition, there are two issues to note in Table 2: (1) it does not have a separate state to represent the plan function. Because the separation of the four functions in the autonomic control loop is only a logical one, and a survey  shows that many ACs rarely implement a separate plan function (due to the fact that it is too costly to implement a fully automatic plan function, or the plan and the analyze functions are usually implemented as a whole, that is, a monolithic analyze function). Therefore, for simplicity’s sake, we omit the plan-related states; (2) the S3 state in Table 2 can be further divided into substates such as E1 to En according to different execution costs, where n depends on the concrete control loop.
If the level of an AC is changed in the left-to-right order in accordance with the above inequality, we call it upgrade, otherwise, degrade. Tuning (i.e., upgrading or degrading) the level of an AC is implemented as setting a proper highestallowed state for the corresponding control loop. To simplify the work of the RSpring tuner, we assume that the execution-costs of an AC (i.e., the resource costs related to the states ranging from E1 to En) are adjusted by the AC itself, thus RSpring only switches the states between Si (i∈[0,3]).
We can see from the above description that, the features of ACs are quite different to that of business functions. For controlling resource allocation and consumption, a business function can only be started/stopped at S0, while an AC can be “natively” tuned at the four different autonomic levels, i.e., S0, S1, S2, and S3. Thus, dynamically upgrade/degrade the levels of ACs can be leveraged to tradeoff business functions and ACs for guaranteeing system performance when resources are limited and competed. The following two facts further ensure its possibility and feasibility. First, many ACs, especially those for monitoring and analyzing purposes, can be carried out without real-time constraints . In other words, it is unnecessary to perform these computations at predetermined or fixed intervals, or complete them at a particular point in time. For the mature ACs, if the execute function is stopped, its monitor and analyze function can still work, and the data being collected or preprocessed by these two functions can be further processed when the execute function is started [2, 3, 4]. That is to say, an AC currently with a relatively lower autonomic level can be upgraded by running more functions in the future iteration of the control loop, and thus consumes more system resources. Therefore, it is possible to bring about flexible resource costs with relatively little negative impact on ACs by upgrading or degrading them dynamically. Second, the varying workloads, usually experienced by a JEE system, support the dynamic execution of ACs. When the workload is relatively low or within an acceptable range, it is reasonable to start running the ACs and upgrade them to make full use of resources. On the contrary, when the workload stays very high, it is feasible to stop the noncritical ACs (e.g., the RTM in Sect. 2.2) or degrade them to save resources for processing more business requests. Of course, the running of those critical ACs (e.g., the one for self-healing and self-protecting) should be guaranteed even if the total available resources are limited. We will describe how to deal with such a case in detail in the next section.
3.2 The gain of an adaptive computation
To tune multiple ACs, the relative importance and cost-effectiveness of them should first be identified. We define the gain of an AC to be the positive impacts with regard to system performance, which are brought by the execution of an AC under a given amount of resources. We leverage the gain values to compare different ACs. What should be noted is that: The gain of an AC is not limited to system performance, but may include other benefits such as security enhancing or reliability improving. However, since the basic goal of our work is to guarantee performance when resources are limited and competed, we consider only the gains regarding performance. The gain value of an AC is directly proportional to the positive impacts (e.g., throughput increasing) brought by this AC for business function processing, and is inversely proportional to the side effects for the correct execution of the whole system (e.g., error increasing). In our model, these values are only used to compare different ACs for tuning the most suitable one, but not for exactly evaluating an AC.
Generally speaking, an AC with a higher gain should be degraded, in frequency and extent, less than that with a lower gain. Of course, there is such a possibility that an AC never wants its level to be tuned (e.g., the critical one for fault tolerance) even if it already has a very high gain value. In such cases, our tuning model allows an AC to specify a pinnedLevel Si (i∈[0,3]). Once reaching the pinnedLevel, any tuning operations on the AC will have no effect. Our model also defines the minLevel Si (i∈[0,3]) due to that some ACs can only be degraded to a specific lowest level. Therefore, if such an AC is degraded, its level cannot be lower than the minLevel. In addition, the importance of an AC to the system performance often relies on the states of the managed objects, and thus changes at runtime. For instance, garbage collection  will become much more important and produce greater gains when free memory is unavailable, and fault tolerance should be applied immediately if a critical error happened. Therefore, to prevent bringing unexpected negative effects to an AC being tuned, our model introduces the concept of SafeRange. We assume that each managed object has an associated SafeRange that can be expressed as something like: “the ratio between the free threads and the total threads in the thread-pool should be higher than 5 %.” If an AC is found that the current state of its managed object exceeds the SafeRange, RSpring will stop tuning the AC for some time, and thus the AC can upgrade or degrade by itself to get back to the SafeRange.
Our work supports two ways to obtain the gain values: (1) Automatic. To be specific, the automatic ways to obtain the gain values are many. One often used is profiling. Online profiling can obtain dynamic gain values, but its running cost is very expensive. Offline profiling is more suitable in the context our work targeted. However, it still requires much preparatory work such as probe-like instrument. What should be noted is that, RSpring leverages only the results of profiling (i.e., a sequence of ACs with comparable gain values). Thus the profiling itself is not our focus. In addition, our work has another way to assign the gain values based on an AC’s maturity  (which implies the importance and cost-effectiveness of the AC to the whole system). For instance, if an AC has all the four functions of the autonomic control loop, while another AC has only the monitor function, the former will be assigned a bigger gain value than the latter. (2) Manual. It requires administrators to assign the gain values, because they are the best candidate to know which AC can contribute most to system performance and to what extent. Such labor-based work is not much and not difficult, because the gain values are used for comparison, which can be assigned approximately.
When tuning multiple ACs, their relationships should also be considered. There are generally two kinds of relationships among ACs. The first is that an AC p uses the managed object of AC q. If q is degraded, its managed object may be cleaned or closed without notifying p, which may cause exceptions in p. In such cases, pinnedLevel and minLevel can be assigned to prevent q from being tuned. The other relationship is that an AC p uses the autonomic functions of AC q, e.g., monitor. If the monitor function of q is stopped, p will be affected. Although in our previous work , we find that there is no such function-sharing cases in the investigated JEE servers, e.g., PkuAS and JOnAS, the tuning model uses the following strategy to address this issue. That is, when an AC is tuned and the tuning requires stopping an autonomic function that is shared by another AC, the level of the formal AC is just degraded without stopping the shared function. In a JEE server, the above two relationships among ACs are often well documented in system manuals such as the JOnAS v5 configuration guide . Therefore, they can be identified with relatively little effort by referring the manuals.
4 The RSpring tuner
AC tuning is similar to process scheduling, but they have something in difference. First, process scheduling is to select a process to consume resources (e.g., CPU), while AC tuning is to degrade (i.e., return some of the possessed resources) or upgrade an AC (i.e., consume relatively more resources). Second, a process has only the states related to stop/start, while an AC may have more states (i.e., S0, S1, S2, and S3). Thus, the process scheduling algorithms are not readily applicable in our work.
Although some relatively complex AC tuning algorithms may be available, e.g., using feedbacks or Markov prediction, they usually cost much more runtime resources . In a resource limited and competed environment, these resource-hungry algorithms should not be used, otherwise the whole system performance may become even worse [25, 26]. We have had tested the Markov prediction scheduling algorithm in a trial experiment. The results are not good enough because the runtime overhead brought about by prediction will offset the tuning effects. In addition, by referring to process scheduling, we find that all the scheduling algorithms are simple and effective. Our tuning algorithms are also inspired by them, which are representative, effective, and have a low overhead.
The mechanism behind these algorithms is the same: enable a flexible tradeoff between business functions and ACs by executing the latter dynamically when resources are limited and competed.
4.1 Two concerns in algorithm design
The essence of AC-tuning is a scheduling problem. In such a problem, priority is always the primary concern. For instance, the priority in the First-In-First-Out (FIFO) scheduling is the time of object-arriving. Even in the Round Robin scheduling, the priority of each scheduled object is implicitly considered (i.e., the priority values are the same). Therefore, the priority of each AC is also the primary concern in our tuning algorithms, and we use the gain value of each AC as its priority of tuning. Additionally, by considering the features of an AC (Sect. 3.1), the priority concern can be manifested at two representative granularity levels: the single state of AC (i.e., fine-grained) and the AC as a whole (i.e., coarse-grained). For instance, there are two mature ACs: L and H, and the gain of L is lower than that of H. “priority highlighted at AC state” means if L is degraded from S3 to S2 (i.e., in fine-grained) before H, then it will be upgraded from S2 to S3after H; “priority highlighted at AC” means if L is degraded from S3 to S0 (i.e., in coarse-grained) before H, then it will be upgraded from S0 to S3after H, although it may be upgraded from S0 to S1before H.
The second important concern is fairness. How to manifest priority upon fairness is challenging. Also, use L and H as an example, if considering only priority, L will be degraded more and upgraded less than H. However, as described in Sect. 3.2, the importance and cost-effectiveness of L may change at runtime. Although, the pinned-Level, minLevel, and SafeRange can be used to address the dynamic gain issue, they cannot prevent the situation that: L will always be degraded while H will always be upgraded. If we consider such a case is unfavorable, then how to apply fairness to prevent it happening? Similarly, by considering the features of an AC, our tuning algorithms apply fairness at two representative granularity levels: the single state of AC and the AC as a whole. Use L and H as an example again: “fairness highlighted at AC state” means if L is degraded from S3 to S2 (i.e., in fine-grained) before H, then it will be upgraded from S2 to S3before H; “fairness highlighted at AC” means if L is degraded from S3 to S0 (i.e., in coarse-grained) before H, then it will be upgraded from S0 to S3before H, although it may be upgraded from S0 to S1after H.
4.2 Algorithms in detail
4.2.1 Covered load-amortizing tune (CLAT): the default algorithm of RSpring
Controlled by RSpring, an AC with a high gain value is upgraded before and degraded after those ACs with relatively low gain values.
If an AC is upgraded by RSpring, the new level of this AC must be lower than or at most equal to the lowest level of those ACs with higher gain values; and if an AC is degraded by RSpring, the new level of this AC must be higher than or at least equal to the highest level of those ACs with lower gain values.
RSpring must not upgrade/degrade the same AC again until it upgrades/degrades all the other ACs. This feature will not work when the SafeRange constraint of an AC is violated (Sect. 3.2) and meanwhile the results of the self level-changing conflict with the above two features.
- Data structure:
- Two double-ended queues (abbreviated to deque) that are used to contain ACs:
remainDeque: contains the ACs that have not been tuned since the last time the algorithm is executed.
changedDeque: contains the ACs that have been tuned since the last time the algorithm is executed.
A list of ACs with their current levels are all in S0 (see Sect. 3.1). These ACs have been sorted in descending order according to each AC’s gain value. All these ACs are copied to the remainDeque in sequence when initializing RSpring.
A system load metric (e.g., CPU utilization) and the corresponding thresholds: lowerBound and upperBound (e.g., CPU utilization between 75 % and 85 %).
- Helper methods:
getCurLoad (metric): get the current system load measured by the given metric.
calibrateACPos (tuningOperation): adjust the position of each AC in between the remainDeque and the changedDeque according to an AC’s current level, gain value, and the tuning operation to be performed. This method is used to guarantee the tree tuning features of the algorithm. As described in Sect. 3.2, an AC can upgrade or degrade by itself without the control of RSpring when the state of the managed object exceeds its SafeRange, and thus may interfere with the three features. Therefore, calibration is necessary.
inSafeRange (AC): given an AC, check if the state of the managed object is in its SafeRange.
initialized (AC): check if an AC has been set to its initial level.
tuning (AC, tuningOp): upgrade, degrade, or initialize an AC.
swap (Deque a, Deque b): swap the given two deques.
The CLAT algorithm is briefly explained as follows: Line 1: get the current system load (curLoad) according to a given load metric (e.g., CPU utilization). Line 2∼21: if the curLoad is lower than the predefined lowerBound, the algorithm will upgrade ACs in accordance with the three tuning features described above. Line 22∼38: if the curLoad is higher than the predefined upperBound, the algorithm will degrade ACs in accordance with the three tuning features. The complexity of the CLAT algorithm is O(n), where n is the number of ACs being tuned. Thus, it is affordable to run in a resource limited and competed environment (the runtime cost of CLAT will be shown in Sect. 5).
The autonomic-level changing diagram of CLAT is shown in Fig. 3(a). In this figure, H and L are both the so-called mature ACs. H’s gain value is higher than L’s. We can see that the curve-changing patterns reflect the three tuning features of CLAT described above. In CLAT, “priority” is highlighted at the granularity of “AC state.” Because the curve of H covers the curve of L in every change (i.e., upgrade/degrade) of AC state. In addition, fairness is also manifested in CLAT due to H has to be degraded before L being degraded once again. That is why CLAT is named as: Covered Load-Amortizing Tune.
4.2.2 Covered carry-through tune (CCTT)
Controlled by the tuner, an AC with a high gain value is upgraded before and degraded after those ACs with relatively low gain values.
Before an AC being upgraded by the tuner, each of the other ACs with higher gain values must have been upgraded to its highest level (e.g., S3), and all the other ACs with lower or equal gain values must not be tuned until the upgrading of the current AC to its highest level is finished; before an AC being degraded by the tuner, each of the other ACs with lower gain values must have been degraded to its lowest level (e.g., S0), and all the other ACs with higher or equal gain values must not be tuned until the degrading of the current AC to its lowest level is finished. This feature will not work when the SafeRange constraint of an AC is violated and meanwhile the results of the self level-changing conflict with the first feature.
The details of the CCTT Algorithm are as follows
isAtHighestLevel (AC): check if an AC has been set to its highest autonomic level.
isAtLowestLevel (AC): check if an AC has been set to its lowest autonomic level.
The differences between CCTT and CLAT are: when upgrading, CCTT will explicitly check if the current AC being tuned is at its highest adaptive level (Line 12). If the check returns false, the current AC will prepare for the next round tuning by being pushed back to remainDeque rather than always being added to changedDeque as in CLAT. Similar characteristics can be found in Line 28 when degrading is performed.
The autonomic-level changing diagram of CCTT is shown in Fig. 3(b). We can see that the curve-changing patterns of H and L in this figure reflect the two tuning features of CCTT described above. We can also see that in CCTT, “priority” is highlighted at the granularity of “AC.” Because the curve of H remains at its highest level S3 when the curve of L degrading from S3 to S0, and because L can only be upgraded until H has been upgraded to its highest level (i.e., carry through). That is why CCTT is named as: Covered Carry-Through Tune.
4.2.3 Other algorithms
RSpring has implemented several other algorithms. For instance, the NLAT (Noncovered Load-Amortizing Tune) shown in Fig. 3(c) is similar to the CLAT algorithm in Fig. 3(a). However, we can see that in NLAT, “fairness” is more highlighted than that in CLAT. Because the curve of L can be upgraded before the curve of H (i.e., the curve of H does not always cover that of L). That is why NLAT is named as: Noncovered Load-Amortizing Tune. For another instance, the NCTT (Noncovered Carry-Through Tune) in Fig. 3(d) is similar to CCTT in Fig. 3(b). Their main difference is that in NCTT, “fairness” is highlighted at the granularity of “AC.” Because in NCTT the curve of L can be upgraded before the curve of H, and because L will be upgraded all the way to its highest level before H can be upgraded.
The algorithms in Fig. 3 are all representative ones to decide when and to what extent an AC can be tuned. However, it is difficult to tell which algorithm is always better than others (see the experiments in Sect. 5). For instance, when the upper bound of the system load is exceeded, whether RSpring should stop an AC immediately and completely (similar to CCTT and NCTT), or degrade two or more ACs in sequence (similar to CLAT and NLAT). This is an open issue. The reason behind is that the runtime effects of a tuning algorithm is affected by many factors, e.g., the system workload and the current autonomic-level of each AC. The detailed comparison of these algorithms is not the focus of this article and will be addressed in the future work.
4.3 The implementation of RSpring
PkuAS is built upon the JMX specification . Its ACs are a set of MBean services. Thus, we implement RSpring also as an MBean service of PkuAS, so that it can get the states of the ACs and control their behaviors through MBean interfaces. JOnAS v5 is built upon JMX and OSGi . Similarly, RSpring is implemented as a service of JOnAS. The other built-in services of JOnAS take the form of OSGi service bundles, which can be regarded as the ACs in our work. The stop/start states of a service bundle can be viewed as the S0 and the Si (i∈[1,3]) level of an AC, respectively. For instance, RSpring uses the stop/start methods of the “org.ow2.jonas.discovery.base.DiscoveryServiceImplMBean” interface to tune the discovery service of JOnAS at the level of S0. The activation and deactivation of the OSGi/JMX methods supplied in such a bundle help to refine the autonomic levels into the specific S1, S2, or S3. For instance, the “disableEJBLogger” and “enableEJBLogger” methods of “org.ow2.jonas.audit.logger.AuditLogComponentMBean” can be used to tune the audit service of JOnAS from S0 to S1 and back forth.
In practice, we find that CPU/memory are the most critical resources in competition between business functions and ACs in most cases [6, 18, 20, 23]. The utilization of these resources can be used as indicators of system loads. Therefore, at present, RSpring supports CPU/memory utilization as a system load metric.
RSpring has its configuration file for PkuAS and JOnAS respectively. To make it work, several parameters should be assigned such as the controlled classes of the AC-compliant JMX/OSGi services identified by administrators, the specific methods in these classes to tune the corresponding ACs at different autonomic levels, and the lower and upper bounds for system load metric. The upper bound should take into account the unpredictable burst of business requests. The gap between the lower and the upper bound should not be too large. Otherwise, very few ACs could be carried out simultaneously. On the other hand, to avoid CPU or memory thrashing, the gap should not be too small either. Therefore, to assign appropriate thresholds, the experience of a system administrator plays an important role, and should take into account the resources hold by the whole JEE system. Some trials are also necessary before assigning these parameters. The newest version of RSpring can be downloaded at the website of PkuAS.
We setup two experiments to evaluate the effectiveness of RSpring by using industry standard JEE benchmark applications and autonomic JEE servers. The first experiment runs ECperf in PkuAS, and uses RSpring to tune the ACs of PkuAS described in the motivating example. The second runs RUBiS in JOnAS, and uses RSpring to improve the system performance by tuning the AC-compliant OSGi services of JOnAS.
5.1 ECperf + PkuAS
5.1.1 Experimental setup
The ACs of PkuAS used in this experiment are RTM, LP, and TPA, which have been described in Sect. 2. Their assigned gain values are 10, 20, and 30, respectively. RTM is assigned the lowest gain value, because it produces the lowest gains with regard to system performance comparing with the other two. RTM has only a monitor function, so its possible autonomic levels are S0 and S1. LP’s possible levels are S0, S1, S2, and S3. TPA is assigned with the highest gain value. Thus, it is guaranteed to run or upgrade first if resources are available to do so. The levels of TPA are also S0, S1, S2, and S3. TPA has defined its SafeRange as below: the current busy threads should not exceed 95 % of all the threads in the thread-pool. If this constraint is violated, RSpring will stop tuning TPA for some time. Thus, TPA can upgrade or degrade by itself to get back to the SafeRange via adjusting the thread-pool size.
The experiment setup remains unchanged as described in Sect. 2. The tested txRate of ECperf is from 3 to 10 and with the sequence 3, 5, 7, 9, 10, 8, 6, 4 to simulate heavy and varying workloads. Such a sequence (with the “normal distribution” pattern) can best simulate the most-appeared workloads in enterprise environment . The testing time lasts 80 minutes. Every 10 minutes, the driver issues a new test with a new txRate. We use CPU utilization as the system load metric, because CPU is the most critical resource in the scenario of ECperf. The corresponding thresholds are set to 75 % (lower bound) and 85 % (upper bound) of CPU utilization, respectively.
5.1.2 Performance of ECperf
The frequency distribution of the NewOrder response times is as follows (see the figure inside Fig. 6). If RSpring uses the default CLAT algorithm and when txRate is 6, there are 412 NewOrder transactions with their response times at 0.2 seconds. However, if PkuAS runs with ACs but without RSpring, the number of this kind of transactions is only 167. The average response time with and without RSpring are 0.983 and 1.572, respectively, and the “response time of 90 % NewOrder” are 2.4 and 3.1, respectively. These values indicate that RSpring indeed improves the system performance by tuning ACs.
From Fig. 6, we can also draw the following three conclusions: (1) PkuAS running with ACs can get a better system performance than the “pure” PkuAS when under heavy workloads (e.g., when txRate is above 6). This is because ACs such as TPA take effects to enable PkuAS do self-management for handling extreme cases. (2) PkuAS running with ACs and RSpring together can get an even better system performance when under heavy workloads. Using the “txRate=10” test for example, the response time of PkuAS running with RSpring is 6.9 (corresponding to the “RSpring (using CLAT)” curve). This value is about 47 % of the response time measured in the “pure” PkuAS test, and is 83 % of the time measured in the “ACs Run (no RSpring)” test. Such a performance improvement is due to the dynamic degrading and upgrading of the autonomic levels of the ACs, which is done automatically by RSpring. (3) The test results of RSpring with different tuning algorithms are close to each other. We cannot tell which one is always better than others, but we can tell that they all bring better results than the tests of “ACs Run (no RSpring).” The reason behind is that all the algorithms reflect the same tuning mechanism: enable a flexible tradeoff between business functions and ACs by executing the latter dynamically when resources are limited and competed.
- The curves in Fig. 7(a) belong to the no-violation pattern. They have the following characteristics: (1) the TPA curve always decreases after and increases before the LP and RTM curves, while the RTM curve always behaves in the opposite way; (2) the TPA curve covers both the LP and RTM curves all the time, and the LP curve always covers the RTM curve; (3) in a round of tuning, the TPA curve will not increase/decrease to a new level once more until the other two curves have both increased/decreased. Thus, we see that these characteristics reveal and conform to the three tuning features of the CLAT algorithm in Sect. 4.2.
The TPA curve in Fig. 7(b) shows that TPA has experienced the violation of its SafeRange constraint when txRate is 7 (during the timeline of the 25∼27 and 28∼29 minutes). In such a case, RSpring will stop tuning TPA for a fixed time interval (e.g., 30 seconds). TPA will automatically adjust the size of the thread-pool (i.e., its managed object) for getting back to its SafeRange, and thus it changes the autonomic level directly to S3 by itself as shown in this figure.
5.2 RUBiS + JOnAS
5.2.1 Experimental setup
The experimented AC-compliant OSGi/JMX services of JOnAS v5.2
It models a work manager, which is similar to the TPA of PkuAS. It allows efficient pooling of thread resources and controls over thread usage
S0, S1, S2, S3minLevel: S1
It is used to audit the ejb and web transactions
It periodically cleans up the work directory
S0, S1, S2, S3
It periodically discoveries the JOnAS domain
S0, S1, S2
It periodically checks and reloads resources
S0, S1, S2, S3
It monitors the application’s deployment process
It checks and redeploys applications
S0, S1, S2, S3
In this experiment, we run the RUBiS  benchmark application in JOnAS and use RSpring to tune the ACs shown in Table 3. RUBiS simulates an online auction website. Workload injection in RUBiS is configured mainly by three parameters: request transition table, client number, and session run time. There are two types of transition table: read-only and read-write, which indicate the transitions of client requests such as from product browsing to bidding. The client numbers used for read-only transition are from 100 to 1500, and for read-write transition are from 100 to 700. The session run time is set to 10 minutes for each client number. Three machines are used in the experiment. Their configurations are: Ubuntu 8.04; Intel Core Duo 1.66 GHZ; 1 GB; and 100 Mb/s. One machine runs the test driver, one runs JOnAS v5.2 hosting the EJB3.0 implementation of RUBiS v1.4.3, and the third machine runs the MySQL v5.5 database used by RUBiS. In the experiment, the database contains more than forty thousand items and one hundred thousand customer records. Due to space limitation, the testing results in this section just show RSpring using the default CLAT tuning algorithm. The tests of the other algorithms can be found on the RSpring website, which can get the same conclusions as below.
5.2.2 Performance of RUBiS
From both Figs. 9 and 10, we can also see that although the system performance becomes poorer when under extremely heavy workloads (e.g., when the clients number greater than 1400 in Fig. 9 and greater than 500 in Fig. 10), RSpring can mitigate the effects of deterioration. Therefore, it is effective to run RSpring with JOnAS-like JEE servers when the whole system is executed in a resource limited environment.
6 Related work
Our work, premised on the assumption that resources are limited and competed, supports a flexible tradeoff between ACs and business functions via the dynamic tuning of ACs. It belongs to the self-optimization category of system performance guaranteeing.
IBM WebSphere provides a mechanism to adjust the application execution environment according to a preferred usage profile . Diao et al.  furnishes the Apache server with self-optimizing ability to keep CPU and memory at a desire level by tuning two parameters of the server. Yang et al.  propose a profile-based approach to developing just-in-time scalability for JEE applications. In this approach, expert’s knowledge of application scaling is captured as profiles. Guided by profiles, an AC-like profiling driver automates the scaling of applications to ensure performance and improve resource utilization, e.g., increases or decreases the EJB instances constituting a JEE application. Although these works use ACs to satisfy end user’s needs or achieve a desired system state, they ignore the costs of ACs. Our work can be referred to as a self-optimization mechanism by considering the costs of ACs together.
Several studies on AC have considered performance penalty [22, 28]. For instance, the JVM garbage collector takes performance cost of garbage collection into account by delaying collection or collecting only parts of the unreachable objects at a time. Such kinds of self-optimization are at a low level and inside ACs, while our approach considers self-optimization at a relatively high level and among ACs. These two approaches are complementary and can be combined into the design of JEE servers.
In the research of runtime verification, probes are instrumented into the programs being verified. Such a probe can capture and send back the interested events occurred during the execution of the program. To reduce the cost of dynamic analysis, studies such as  exploit the results of static typestate analyses to reduce the number of probes and the amount of information to be recorded. Thus, we can view the probes as the monitor-level ACs, which can be tuned by RSpring.
A utility based approach has been used for allocating resources among different computations [8, 14]. The main idea is to first use a utility function to express the relationship between high level business goals and system performance, then use an empirical model to correlate system performance with resources allocation, and finally combine these two functions for representing a utility model that guides the allocation of resources. The main difficulty of applying such an approach is to obtain the empirical model that exactly represents the relationship of resource allocation and performance. In addition, the runtime resource costs of the utility functions are usually high . Therefore, our work does not use utility functions for tuning ACs at present, especially when targeting a resource limited environment.
Several studies such as [7, 15, 16] try to guarantee the overall system performance by upgrading and degrading the service levels of business functions. When the system load is heavy, these functions will degrade their service levels for achieving a steady and high throughput, e.g., return the top 100 search results to a client instead of all the results, return the results without sorting, or just decrease the search quality. Studies such as  leverage admission control to guarantee high system performance. The application will refuse to serve any new incoming client requests when the system is under heavy stress. Other studies such as  schedule incoming requests to the order in which concurrent requests can be served. For instance, the application will distinguish the type of client requests and will first serve the ones with the shortest remaining processing time. In this way, more requests can be served. All the above approaches require the applications to be modified or built from the start to have the capability of dynamic service-level negotiation. Our approach changes the levels of ACs instead of business functions, and thus leaves the business logic intact.
7 Conclusion and future work
With the increasing use of autonomic computing technologies, JEE servers are implemented with more and more adaptive computations. It brings challenges to resource allocation for achieving the performance goals because of resource competition between business functions and adaptive computations, especially when the whole system is under heavy load. In this article, we have presented a tuning model for the performance improvement of applications in JEE servers. This model is built on the autonomic control loop, and the features and gain of an adaptive computation have been considered in the model. Based on the model, the autonomic level of an adaptive computation is upgraded or degraded dynamically so as to control its resource cost. We have implemented the RSpring tuner for JEE servers such as PkuAS and JOnAS, which has been evaluated by ECperf and RUBiS benchmark applications respectively. The results show that RSpring can effectively improve the application performance by 13.6 % in PkuAS and 19.2 % in JOnAS with the same amount of resources.
In the future, we plan to build the RSpring tuner for other JEE servers such as JBoss and Glassfish; investigate other suitable tuning algorithms; and try to integrate this work with the work of allocating extra resources, especially in a virtualized environment such as a cloud computing platform.
The authors would like to thank all the anonymous reviewers for their valuable comments and suggestions to improve the quality of the article. This work is supported by the National Basic Research Program of China (973) under Grant Nos. 2009CB320703; the National Natural Science Foundation of China under Grant No. 61121063, 60933003; the High-Tech Research and Development Program of China (863) under Grant No. 2012AA011207; and the IBM University Joint Study Program.