A Practical Method for Estimating Performance Degradation on

Preparing to load PDF file. please wait...

0 of 0
100%
A Practical Method for Estimating Performance Degradation on

Transcript Of A Practical Method for Estimating Performance Degradation on

A Practical Method for Estimating Performance Degradation on Multicore Processors, and its Application to HPC Workloads
Tyler Dwyer, Alexandra Fedorova, Sergey Blagodurov, Mark Roth, Fabien Gaud, Jian Pei School of Computing Science, Simon Fraser University
{tdwyer,fedorova,sergey blagodurov,mroth,fgaud,jpei}@sfu.ca

Abstract—When multiple threads or processes run on a multicore CPU they compete for shared resources, such as caches and memory controllers, and can suffer performance degradation as high as 200%. We design and evaluate a new machine learning model that estimates this degradation online, on previously unseen workloads, and without perturbing the execution.
Our motivation is to help data center and HPC cluster operators effectively use workload consolidation. Data center consolidation is about placing many applications on the same server to maximize hardware utilization. In HPC clusters, processes of the same distributed applications run on the same machine. Consolidation improves hardware utilization, but may sacrifice performance as processes compete for resources. Our model helps determine when consolidation is overly harmful to performance. Our work is the first to apply machine learning to this problem domain, and we report on our experience reaping the advantages of machine learning while navigating around its limitations. We demonstrate how the model can be used to improve performance fidelity and save energy for HPC workloads.
I. INTRODUCTION
Workload consolidation refers to a resource allocation principle where we try to place many applications or processes comprising a distributed application on the same server, so as to not leave any cores idle. Consolidation makes a fundamental trade-off between performance and hardware utilization (and thus power efficiency). We improve utilization of the machine’s resources, but sacrifice some amount of performance, because increased resource sharing among threads may reduce the rate of retired instructions.
Previous studies have shown that performance degradation occurring when threads or processes run on the same multicore CPU and share resources, such as last-level caches, memory controllers, system request queues, and prefetch bandwidth, can reach as much as 200%, relative to running in isolation [1]–[5]. Such severe degradation can defeat the benefits of consolidation, leading to violation of customer QoS constraints or simply producing application slowdowns that are deemed unreasonable [6], [7]. In certain cases, the slowdown could be so extreme, that despite saving power we waste energy, because the workload takes much longer to complete under consolidation than without it [8].
It is clear that we cannot use workload consolidation without considering its potentially damaging effects. Unfortunately, conventional performance tools, while allowing us to observe events like retired instructions or cache and memory controller

accesses, do not tell us how much performance degradation the workload is experiencing. If we have the luxury of knowing in advance the input used by our workload and have access to the hardware on which it will run in the field, we can execute the workload offline with various degrees of consolidation, determine the optimal, and use that setting in the field. Unfortunately, in most cases we do not have advance knowledge of runtime parameters, and so deciding whether or not to consolidate becomes guesswork.
We propose a method that helps making better consolidation decisions in data centers. We develop a model that takes as inputs performance counter values obtained on a consolidated workload (which can be measured inexpensively online) and produces an estimate of how much performance degradation each thread or process is suffering relative to running on the CPU alone, without contention for resources. We do not require running applications in isolation or perturb their execution. Using this estimate, the operator can decide if the degradation is beyond an acceptable threshold, and if it is distribute the workload to a larger number of CPUs or servers. The model is designed for scenarios where either different applications are consolidated on the same machine, or processes of the same application share the hardware (as commonly happens in a HPC setting).
We create the model using machine learning; to the best of our knowledge, our work is the first to apply machine learning to this incredibly complex problem. We train the model offline, using a set of widely available benchmark programs. To maximize accuracy, the training programs should have properties similar to the workload on which the model will be used, but they need not be the same or overlap. For example, if we are targeting business applications, we could train on SPEC JBB or TPC-W; for scientific workloads, we could train on SPEC CPU. The model is trained for specific hardware, but this needs to be done only once, and can be performed as part of system installation and configuration.
Although machine learning has been used to model system behaviour in the past, it has not been applied to performance degradation on multicore systems. We were compelled to try machine learning, because it could help overcome practical limitations of previous solutions. Previous solutions used analytical modeling, heuristics based on hardware counters, or trial-and-error methods. Analytical modeling is extremely

SC12, November 10-16, 2012, Salt Lake City, Utah, USA 978-1-4673-0806-9/12/$31.00 c 2012 IEEE

fragile and challenging, because modern CPUs are incredibly complex; moreover, crucial details about the microarchiecture are often unavailable due to intellectual property protection. Heuristic-based models provide only a coarse approximation of performance degradation; typically they can tell us whether the degradation is occurring, but not its magnitude. Indeed, as we show in Section IV, a cluster scheduler based on a heuristic model can waste energy relative to the scheduler that is based on the more precise model proposed here. Trial-and-error methods require running applications in different co-schedules with other applications [9] or with dummy workloads [7] in an effort to find a co-schedule that minimizes the degradation, but as the number of cores increases the number of co-schedules grows as well, as does the perturbation imposed on the workload. In considering the limitations of previous approaches, machine learning looked like a promising alternative.
Machine learning excels at discovering complex relationships between a variety of factors and filtering out the factors that are not important. This seemed very well suited to our problem, because modern CPUs allow monitoring hundreds of performance events, some of which correlate with sharinginduced degradation, but it is very difficult to filter the spurious events manually. Our idea was to use machine learning to discover the relevant hardware events automatically, so we could apply the same methodology on any hardware platform, to automatically create a model for the desired processor without the manual labour of picking the right counters. Indeed, we observed that our model-building methodology seamlessly ported between two machines with different hardware counters, cache and system bus architectures.
The key limitation of machine learning is that the model is only as good as the training data. While we can make an effort to train on the workloads similar to those used in the field, it is not always possible. In evaluating the model, we observed that if the input data falls outside the range of the values seen in training, the model becomes limited in its ability to make accurate predictions. While it is possible to re-train the model using the “outlier” workload, it is crucial to detect in real-time when the model is about to produce a poor estimate. To that end, we propose a statistical solution, called confidence predictor, which detects when the model is about to make a mistake. Upon detecting the “low confidence” signal, the operator can conservatively decide to not consolidate this workload, or re-train the model using the data obtained on the “outlier” workload. The confidence predictor reduces the maximum error by a factor of 2× or 3× in our experiments, while marking roughly 25% of the predictions as “nonconfident”.
We focus on modeling the degradation resulting from sharing the multicore chip’s resources, because they are the most difficult to control in software, unlike CPU cycles, memory space and disk/network bandwidth, which can be controlled by quotas. In fact, in the environment where we chose to evaluate our model, an HPC cluster running scientific workloads, each thread is typically given a dedicated CPU, and a job’s memory is sized to fit within the physical memory limits of the

machine. Network bandwidth on the cluster interconnect was not a bottleneck in our experiments, but if it were, machine learning could also be used to model contention for that resource.
The contributions of our work are: (1) creating the methodology for modeling performance degradation on multicore systems using machine learning, (2) evaluating the strengths and limitations of the resulting model, (3) designing a confidence predictor that signals when the model is unable to produce an accurate estimate, and (4) demonstrating how the model can be applied to improve performance fidelity and save energy in an HPC setting.
The rest of the paper is organized as follows: Section II describes the methodology for building the model. Section III evaluates its accuracy. Section IV presents and evaluates a simple scheduler for HPC clusters that uses our model. Section V discusses related work. Section VI summarizes our findings.
II. THE MODEL
Before we describe the methodology for creating the model, we explain how the model would be used in practice. We begin with the assumption that our target application is a singlethreaded process. Then we explain how the model would work with multi-threaded processes.
First, the user would train the model following the procedure described below, and using as the training set the applications that most closely resemble those on which the model would be used in the field. Each application is executed alone on the machine and in combination with other applications. We compute how much slower the application runs when coscheduled with others relative to running alone; this slowdown is the performance degradation. We then train the model to predict performance degradation using hardware counter values obtained on the system when the target application runs with other jobs. The resulting model is thus set up to estimate how much slowdown the application is suffering when spacesharing the CPU with others, without the need to execute the target application alone.
The model estimates the degradation based on per-core and system-wide hardware counter values, and so it is agnostic to whether the cores are running single-threaded processes or threads from the same application. We purposefully do not configure the model to account for positive effects of cooperative sharing; such models are available and, if desired, can be used in conjunction with the proposed model [10].
The goal of the model is not to facilitate contention-aware scheduling of threads within an application, but to decide whether we need to allocate more hardware in environments where many applications runs on the same physical server. So if we have a multi-threaded application sharing hardware with other processes, we would use the model to estimate, for each thread, the performance degradation that the thread is suffering under resource contention relative to running alone and then average the degradations of all threads to obtain the degradation for the entire application. If the application is

Cores per socket Shared per socket

Intel: 2-socket “Clovertown”
4
Two L2 caches (per pair of cores), front-side bus, prefetcher, memory controller

AMD: 2-socket “Istanbul”
6
L3 cache (all cores), system request queue, memory controller, data and memory controller prefetchers

TABLE I SHARED RESOURCES IN THE EXPERIMENTAL SYSTEMS

deemed to suffer unreasonable performance penalty, we would migrate that application to a less loaded server, to create a less contentious environment. Within each server, a local OS or hypervisor scheduler, ideally one that takes into account both resource contention [5] and co-operative resource sharing [10], [11] will decide how to place threads on cores. Other resourceallocation decisions, e.g., regarding CPU and memory quotas, can be applied on top; deciding how to combine allocation of resources of different types is deferred to future work.
In the rest of the section we explain how we build the model. As will become clear in Section II-B2, our model is a decision tree. A decision tree consists of nodes and branches, where each node is labeled with an attribute (e.g., a hardware counter type in our model) and a threshold for the attribute’s value. Based on the thresholds we decide which of the branches emanating from the node to follow. We follow the tree all the way down to a leaf, comparing the measured hardware counter values with the corresponding thresholds assigned to the nodes. The resulting leaf node will give us the predicted degradation for the data point characterized by these hardware counter values. So the goal of building the model is to assign the right attribute thresholds to intermediate tree nodes and predicted degradation values to the leaves, so we arrive at a reasonably accurate prediction of performance degradation. The process of building the model consists of three steps: (1) collection of the training data, (2) attribute selection, (3) model training.
A. Collection of the training data
1) Testing Platform: To confirm portability of our methodology, we built and tested our model on two systems, Intel and AMD using exactly the same procedure. Refer to Table I for system parameters. We trained and tested the model only on a single socket, because additional contention from running applications on the second socket did not significantly affect degradation [5]. The (AMD) system has a NUMA (nonuniform memory access) architecture, and so we ensured that an application’s memory is allocated on the same node as where the application runs. This is how the OS would typically behave1.
2) Applications and Co-schedules: Since our goal was to evaluate the model on scientific applications typical of HPC clusters, we trained the model on the SPEC CPU2006 suite.
1On most recent NUMA systems with directory-based cache coherence protocol, processes running on separate memory nodes and sourcing data from their local memory node will not noticeably affect each other’s performance

We ran applications in three different types of co-schedules. The solo run is when an application runs alone on a system without contention. The solo run was used in calculating the true value of performance degradation. True degradation is needed only to train the model and to evaluate the accuracy of predictions; we do not expect to know it on a production system. In the other two types of co-schedules the application runs in contention, concurrently with others. In the clean coschedule, the primary application (the one whose degradation we predict) runs with several copies of itself (three on the Intel system, five on the AMD system). In the random coschedule, the primary runs with randomly chosen benchmarks on the remaining cores.
To collect the data for training, we start all applications in the co-schedule at the same time. If any interfering application terminates before the primary is finished, we restart the interfering application, thus keeping the primary in full contention for the entire run. Overall, we recorded over 10 random coschedules for each of the 27 applications. Together with clean and solo runs, this resulted in over 500 co-schedules.
3) Recording Performance Events: To record the attributes relevant for modeling degradation, we use the hardware performance counters that can be used to measure events, such as last-level cache misses, the number of bus transactions, etc. (some shown in Table III). On the Intel system there are 340 event counters per core, but only four hardware registers for counting them. To be able to record all these events, we had to sample them by switching between different event types.
When we switch between events, we have to make sure that each event is sampled for a substantial period of time, to ensure good sampling accuracy. The need to measure a large number of events puts a lower bound on the interval of execution for which we are able to record all available counters. We set that interval to 5 billion instructions, which allowed us to capture the required events without major loss of precision2. Each 5-billion instruction window is called an execution instance. We train the model and produce predictions for execution instances, as opposed to the entire program. The implication is that in production setting, we need to sample event counters for 5 billion instructions (a few seconds of execution time) before we are able to produce an estimate of the degradation. Therefore, the model is best suited for long-running workloads, such as the scientific applications we evaluate in our study.
As we rapidly switch between different types of counters, intermittent system events, such as handling of an interrupt, can introduce unexpected spikes or dips in the measurements. Data containing these variations presents a challenge for a machine learning algorithm, because there is not enough training data to learn the behaviour during these extraneous events. To smooth out their influence, we found it helpful to represent attribute values for each instance as the rolling average of
2Adding the attribute selection step, described below, enabled us to use a smaller execution interval. We observed, however, that the accuracy of the model was not sensitive to the size of that interval.

the past ten instances, with an exponential preference to more recent instances.
4) Calculating performance degradation: The degradation for an instance is obtained by calculating the percent increase in clock cycles needed to complete the fixed instruction window under contention, relative to solo. For example if the co-schedule A-B-C-D has 50 execution instances, then the A solo run would also have 50 instances, as instances are based upon retired instructions, which are constant regardless of contention. The clock cycles of the ith instance of A-B-CD, denoted ABCDcilk, are compared with the ith instance of A’s solo run, denoted as Aiclk. From these, the degradation of A while in contention with B-C-D is computed as:

Deg(Ai ) = ABCDcilk − Aiclk ∗ 100%

(1)

BCD

Aiclk

We performed the above calculation on all instances in our data set. After this procedure our dataset contains equally-sized execution instances, each with 340 × 4 = 1360 attributes from the event counters, and the degradation value.

B. Attribute Selection and Model Training
1) Attribute Selection: Before building the model we reduced the number of attributes in the dataset from 340 per core to 19 per core. Attribute selection was performed for three reasons. The first is to eliminate the attributes that were redundant or unrelated to degradation. The second is to reduce the training time from up to several hours (with 340 attributes) to several minutes (with 19 attributes). Third is to allow a new dataset to be recorded with fewer events sets, leading to more accurate recording.
We performed attribute selection using a suite of machine learning algorithms Weka. We tested several attribute selection techniques by creating models for each set of selected attributes and comparing their accuracy. The technique with the lowest error rate was correlation based feature subset attribute selection (CfsSubset) [12]. CfsSubset sorts the attributes by their correlation to the class attribute (degradation) and to the other attributes in the dataset. Attributes with a high correlation to the class attribute and a low correlation to the other attributes are considered relevant.
Attribute selection yielded 19 attributes for each core, shown in Table III; same 19 attributes on each core were selected. In order to train the model, we want to distinguish the counters of the target core (the one whose degradation we are predicting) from those of the interfering cores. To do that, for each event counter we average the values obtained on all interfering cores, for a given execution instance. In summary, before we train the model, we have thousands of execution instances, and for each we have the values of the event counters on the target core, the average of the counters on the interfering cores, and the degradation.
2) Model Creation: In the process of building the model we evaluated all modeling algorithms available in Weka, listed in Table II. Although we are unable to provide detailed analysis of all these methods due to space constraints, we settled on

Bagged REPTree Gaussian Process REP Tree Isotonic Regression SVM Reg SMO Reg M5P Pace Regression

+0.00% +1.20% +2.35% +3.79% +3.84% +3.85% +4.14% +4.36%

Linear Regression PLS Classifier Decision Table Simple Linear Regression Neural Network Conjunctive Rule Decision Stump M5Rules

+4.45% +4.76% +6.12% +8.10% +10.60% +11.17% +12.18% +17.84%

TABLE II MODELLING TECHNIQUES TESTED AND THE AVERAGE ERROR THEY PRODUCED RELATIVE TO REPTREE. REPTREE’S AVERAGE ERROR WAS
16% IN CROSS-VALIDATION.

REPTree, because it yielded the highest accuracy in a variety of test cases. REPTree can be used in two modes: as a regression tree, where the predicted outcome a real number, and as a classification tree, where the predicted outcome is a class, or a range of degradation values in our case. Regression mode produced a higher accuracy than the classification mode, so we use it in our model.
We experimented with several accuracy-improving techniques and found bagging to be very effective. Bagging, also referred to as bootstrap aggregating, is known to lower the error rate, reduce the variance and help avoid over-fitting. Bagging works by creating m new datasets each populated through sampling from the original dataset, uniformly with replacement. From each of these m data sets a new model is created, providing us with m models. Each model is then used to produce an estimate, and all these estimates are averaged to create the final estimate. When bagging was used, the average error of our REPTree model reduced by over 6%, and the time to train the model and make predictions remained reasonable.
The REPTree model is well suited for online use, because an estimate can be produced exceptionally quickly (tens of microseconds in our experiments). The tree is represented as a table, and all that is required is a few table look-ups, whose number is proportional to the depth of the tree.
III. RESULTS
We begin by taking a closer look at the decision tree created for the Intel system (Section III-A). Then, in Section III-B we present the quantitative evaluation of the model’s accuracy, along with the analysis of scenarios producing high errors, and the confidence predictor, a solution for anticipating inaccurate predictions online. We are unable to present detailed results for the AMD system, but the conclusions we reached from that data were qualitatively similar.
A. Analysis of the decision tree
A decision tree is structured as a collection on nodes and branches, where each node contains the attribute used for making the branching decision and the corresponding threshold value. For instance, in our tree for the Intel system the attribute showing the number of delayed bus transactions (L2 REJECT BUSQ:MESI from Table III) is used for branching at the root of the tree. Instances that generated fewer than

Event Name UNHALTED CORE CYCLES RS UOPS DISPATCHED CYCLES:PORT 1
SSE PRE EXEC:NTA L2 M LINES OUT:BOTH CORES
L2 REJECT BUSQ:MESI
L1D CACHE LOCK:M STATE BUS TRANS INVAL:BOTH CORES
BR MISSP EXEC SIMD UOP TYPE EXEC:SHIFT SIMD COMP INST RETIRED

Event Description
Clock cycles elapsed.
The number of cycles for which micro-ops are dispatched for execution on port 1. Indicative of processor utilization.
This is a software prefetching event. Counts the number of times the SSE prefetch NTA instruction is executed.
Counts the number of L2 modified cache lines evicted by both cores. Indicative of the pressure on the memory hierarchy.
Counts event when a pending L2 cache request that requires a bus transaction is delayed from moving to the bus queue. This can happen, for instance, when the bus queue is full.
Counts the number of locked data reads in modified state from cacheable memory.
Counts invalidate bus transactions for both cores, which can be generated, for instance, when a cache line write misses the L2 cache.
Counts the number of mispredicted branch instructions.
SIMD packed shift micro-ops executed.
Retired computational Streaming SIMD Extensions (SSE) scalar-single instructions.

Event Name INSTRUCTIONS RETIRED SEGMENT REG LOADS

Event Description The number of retired instructions. Number of segment register loads.

SSE PRE EXEC:STORES L2 RQSTS:M STATE L1D CACHE ST:I STATE

This is a software prefetching event. Counts the number of times SSE non-temporal store instructions are executed.
Counts all completed L2 cache requests, including hardware prefetches. M STATE counts accesses to cache lines whose content differs from that in in the memory).
Counts the number of data writes to cacheable memory that missed the cache.

LOAD HIT PRE BUS TRANS P:ALL AGENTS

Counts load operations conflicting with a software prefetch to the same address.
Counts all partial bus transactions.

BR IND MISSP EXEC INST RETIRED:STORES

Counts the number of mispredicted indirect branch instructions.
Counts the number of instructions retired that contain a store operation.

TABLE III LIST OF THE ATTRIBUTES SELECTED BY ATTRIBUTE SELECTION FOR THE INTEL SYSTEM

5 delayed bus transactions per thousand instructions follow the right branch, the rest of the instances follow the left branch.
Examining the Intel tree, we observed that the number of delayed bus transactions was one of the most important attributes in the decision tree, as it is used for branching decisions at almost every tree level starting from the root and appears as the branching attribute in 10% of all the tree nodes. A delayed bus transaction indicates that a memory request has to wait in the queue before it is issued. Frequent occurrence of this event is indicative of memory bus contention. Earlier work studying contention on the same Intel system observed that memory bus contention was the key cause of performance degradation [5]. Our model identified this event as important automatically. Other events that are used most frequently in decision making include L1 store misses (L1D CACHE ST), L2 cache requests (L2 RQSTS), evicted L2 cache lines that were dirty (L2 M LINES OUT) and other events indicative of the bus traffic (BUS TRANS INVAL and BUS TRANS P).
Somewhat surprisingly, we observed that many of the most significant attributes were related to the write intensity of the workload, e.g., L1D CACHE ST and L2 M LINES OUT. This could indicate that the underlying system is not able to buffer the writes to the extent that their effect on the memory system is minimal.
The strongest positive correlation with degradation was observed for the following attributes: L2 REJECT BUSQ (0.87 correlation coefficient), UNHALTED CORE CYCLES (0.64), BUS TRANS P (0.48), L2 M LINES OUT (0.26) and L1D CACHE ST (0.22). This makes sense, as these attributes are indicative of the memory bus contention: UNHALTED CORE CYCLES (the number of cycles it took to execute the instruction window) will be high for applications that experience large memory-bus delays, as will be the number of bus transactions (BUS TRANS P), the number of evicted last-level cache lines (L2 M LINES OUT) and the number of stores (L1D CACHE ST). The strongest negative

Fig. 1. Cumulative distribution of errors
correlations were observed for INST RETIRED.STORES (0.25) and BR IND MISSP EXEC (-0.21).
B. Evaluation of the model’s accuracy on the Intel system
We evaluate the model’s accuracy using cross-validation. For each primary benchmark, we predict its degradation in the clean and random co-schedules. To produce the estimates, we first remove from the data set all execution instances containing this benchmark (either as the primary or interfering). We train the model on the reduced data set and then produce the estimate of the degradation. This way we ensure that the model is not trained on any instances of the application for which it is trying to make predictions. This is the most rigorous validation procedure of all the available options.
Our metric of accuracy, the error rate, is the absolute difference between the estimated and the actual measured degradation. For instance, if the measured degradation was 5%, but we predicted 7%, the error would be 2%. Figure 1 shows the cumulative distribution of prediction errors produced by our model. We observe that 80% of the errors are under 20%. The average error was 16%.
Figure 2 shows the estimated and predicted degradation over time (for all instances) for two benchmarks that are representative of the results that we observed. Each benchmark was run in many co-schedules; since we are unable to show

Fig. 2. Estimated vs. predicted degradation for all instances over time for two selected benchmarks. We show the co-schedules that produced the lowest, median and the highest errors.

all co-schedules due to space limitations, we report the data for the co-schedule that produced the smallest prediction error (Min Error Co.), median error (Median Error Co.) and the highest error (Max Error Co.). In the case of tonto we observe that the model is very good at following degradation trends over time; we observed similar behaviour with other benchmarks exhibiting temporal variation in degradation.
Since we are unable to show detailed time-series graphs for every benchmark, we present a summary in Figure 3. The x-axis shows the degradation values, the y-axis shows the primary benchmarks. For each benchmark, the dot indicates the true degradation, and the triangle indicates the value predicted by the model. The length of the line connecting the two symbols correlates with the magnitude of the error. The first chart on the left shows the results for the co-schedules that produced the smallest error for each benchmark, the second chart shows the median-error co-schedules, the third chart shows the highest-error co-schedules.
Min- and median-error charts show that, with a few exceptions, the prediction errors are quite small. From the highesterror chart we observe that there are a few large errors for high-degradation benchmarks, such as lbm, soplex, libquantum and mcf. As we will demonstrate in the next section, these benchmarks show behaviour that is distinct from the other benchmarks in the training set. Since cross-validation ensures that we do not train on the benchmarks whose degradation we are trying to predict, the model is not trained to recognize these “outliers”. The fourth chart in Figure 3 shows how the results for highest-error co-schedules improve when we apply the confidence predictor; these data and the confidence predictor will be explained in Section III-D.
C. Uncovering the outliers
With cross-validation, the training set contains absolutely no instances of the application whose degradation we are trying to predict. So, for instance, if we are predicting lbm the model had no event counters involving lbm in its training set. Therefore, if lbm’s attributes are very different from the training set, we could use this variation to explain inaccurate predictions and, furthermore, to anticipate them. To test this theory, we compare performance attributes of the outliers to those in the rest of the set.
Figure 4 compares the event counter values of four benchmarks with the highest errors in the max-error coschedules (lbm, libquantum, mcf and soplex) with those of the other benchmarks. We begin by analyzing lbm. Figure 4 clearly shows that lbm is an outlier in the attribute L2 M LINES OUT. This event count is high when the work-

load evicts a lot of L2 cache lines in the modified state. When the evicted cache lines are dirty, the memory system has to do more work in handling cache misses, because modified lines must be written back to memory. Lbm happens to be a very write-intensive application [13], and since its value of L2 M LINES OUT is extremely high, our model extrapolates an extremely high degradation. In reality, writes, while certainly adding pressure to the memory system, contribute to execution latency only indirectly, since they are handled asynchronously. (The correlation of L2 M LINES OUT with performance degradation is only 0.26). As a result, our model greatly overestimates the degradation.
We now look at libquantum. In sharp contrast to lbm, libquantum performs very few writes [13] and is unique in its extreme latency sensitivity. It has very poor cache reuse and as a result spends 100% of execution time in memory episodes, where it is waiting for at least one memory request [14]. Indeed, in Figure 4 we observe that libquantum has an unusually high value of the attribute L2 REJECT BUSQ:MESI, which occurs when a pending data request from the L2 cache is delayed from moving to the bus queue, and is indicative of long memory episodes. No other benchmark in the SPEC CPU2006 suite has similar behaviour. As a result of its latency sensitivity, libquantum suffers significantly more from contention than other benchmarks, because any increase in memory-system latency has a direct effect on its performance. Since libquantum is the only benchmark with this behaviour, the model is unable to capture its extreme latency sensitivity and so it consistently underestimates the degradation for libquantum.
Looking further at soplex, we observe that it has a vastly different prefetching behavior than the rest of the benchmarks. First of all, it has a somewhat higher count of software prefetch events (SSE PRE EXEC:NTA) than other benchmarks. But what is particularly interesting is that it has a dramatically high count of LOAD HIT PRE events, which counts load operations conflicting with a software prefetch to the same address. For every other benchmark in the suite except one3, this event count is close to zero. This means that soplex performs extremely effective prefetching, since a large number of loads already have a corresponding prefetch request in flight. Effective prefetching masks memory latency, and so contention for shared resources has a much smaller impact on soplex’s performance than one might expect. Since soplex is the only benchmark with such property, the model is unable
3Gobmk also has a high occurrence of this event, but it is still roughly four times smaller than for soplex.

Primary Benchmark

416.gamess 454.calculix 400.perlbench 453.povray 456.hmmer
444.namd 465.tonto 464.h264ref 447.dealII
481.wrf 434.zeusmp 445.gobmk 437.leslie3d 410.bwaves
458.sjeng 435.gromacs
473.astar 401.bzip2 459.GemsFDTD 436.cactusADM
433.milc 403.gcc 471.omnetpp 429.mcf 462.libquantum 450.soplex 470.lbm

Min Error Co.
q q q
q q q q
q q
q q q q
q q q
q q
q q
q q q q q q q

Median Error Co.
q q
q q
q q
q q
q q
q q
q q q q q q q q
q q
q q
q q
q

Max Error Co.

q q q q q q q q q q q q q q q q q q q q q
q q
q q

q q

Conf Max Error Co.

q q q q q q q
q q q
q

Legend q True Predicted
q q
q

q q q
q q q q q

q

0% 50% 100% 0% 50% 100% 150% 0% 50% 100% 150% 200% 250% 300%0% 50% 100% 150%
Degradation Value
Fig. 3. Difference between the actual and predicted degradation for the best, median and worst predicted co-schedules for each primary benchmark. The right-most chart shows the max-error co-scheduled when we apply the confidence predictor.

Fig. 4. Highlighted event counters for libquantum, lbm, soplex and mcf. (Must be viewed in colour).

to correctly factor in the effect of successful prefetching and so it typically overestimates soplex’s degradation.
Finally, examining the outlier events for mcf, we see that, similarly to lbm and libquantum, it has a large number of L2 REJECT BUSQ and L2 M LINES OUT events. Although when we validate mcf, the training data does include lbm and libquantum, the number of the instances that include these benchmarks is very small, relative to the other instances that have much lower counts of the two events. As a result, the model is not strongly trained to make accurate predictions in this case.
The key insight that we gain from this analysis is that the applications responsible for the highest errors can be predictably identified if we analyze how their performance attributes are different from those of other applications. We use this insight to create a method for anticipating when the model is likely to produce a high error.

D. Confidence Predictor
The confidence predictor decides whether or not the model is likely to make an accurate prediction by comparing the hardware counter attributes of the instance whose degradation we are about to predict with the distribution of attribute values seen in training. If two or more attributes of the to-be-predicted instance are more than two standard deviations away from the mean of the values seen in training, the confidence predictor marks the instance as non-confident and produces a null prediction. There are several ways how the operator can handle null predictions. One possibility is to conservatively label this workload as high-degradation and not to consolidate it – this policy can be used for workloads with strict QoS requirements. Another solution is to add a copy of this workload to a database of training benchmarks, so that the model can be trained on it in the future.
The fourth chart in Figure 3 shows the highest-error co-

schedules once we introduce the confidence predictor. We observe that the magnitude of errors is substantially reduced. As the trade-off, the predictions for a few benchmarks are not made: this would occur if all instances for that benchmark are marked as non-confident. This is expected: if the behaviour of the benchmark is drastically different from the behaviour seen in training, the model would not be able to confidently predict any of its instances. In this case, it is best to either re-train the model on the benchmark or to conservatively assume a high degradation. Overall, about 25% of the instances are omitted as non-confident.
Figure 5 shows the scatter plots of the errors produced by the model when we make the estimates for all instances, regardless of confidence, and when we produce the estimates only for confident instances. The confidence predictor substantially helps to filter out erroneous predictions, reducing the maximum error by about a factor of two. The average error for all instance is around 16%, but it is reduced to 10% when the confidence predictor is applied. On our AMD system, the confidence predictor reduces the maximum error by about a factor of 3× and improves the average error from 13% to 10%, while marking 25% of the instances as non-confident.
In summary, we conclude that machine learning is a reasonable method for estimating the complex effect of sharinginduced performance degradation on multicore processors. It is able to produce accurate predictions in the majority of the cases, but when the training set has insufficient diversity we can anticipate high errors in the model by applying the proposed confidence predictor.
IV. USE CASE FOR THE MODEL In this section we describe how our model could be used for improving resource-allocation decisions in high-performance computing (HPC) clusters. HPC clusters run scientific applications, many of which are structured as multi-process jobs communicating via MPI (message-passing interface). By default, cluster scheduling algorithms, such as Maui [15] or Moab [16], will assign a process of a given job to every available core on the server, but since many MPI applications are very memory-intensive, they will experience substantial
Fig. 5. Errors for all instances (left) and only confident instances (right).

performance degradation when the processes of the same job share a multicore CPU [17]. The cluster operator may want to spread the job’s processes across a larger number of nodes, so as to avoid unreasonable performance degradation, but with existing tools it is difficult to decide whether the degradation is high enough to justify using extra hardware.
To demonstrate how the proposed model can address this problem, we prototype a new cluster scheduler that uses the model for scheduling decisions. The proposed scheduler, described later, improves on two baseline cluster scheduling policies: Best-fit and Min-collocation. Best-fit, which is the most commonly used policy, allocates the processes of the same job on all available cores on the node, using additional nodes if needed, but if a single job does not fill all the cores, it fills them with processes of another job. The other baseline policy, Min-collocation, attempts to schedule no more than one job per node, as long as there are unused nodes available.
The Best-fit policy ensures maximum hardware utilization, but allows contention for multicore resources. Min-collocation would produce less resource contention, but will use more hardware (and power). We demonstrate how to find the balance between these extremes with a new Balanced scheduler that relies on our model.
The Balanced scheduler initially assigns jobs to nodes following the Best-fit policy. It then begins monitoring the hardware counter values selected by the model, and estimates the performance degradation for each job. If the degradation is estimated higher than the acceptable threshold, set to 50% in our experiments, the Balanced scheduler starts up an additional server and migrates the suffering job to that server. The scheduler is also able to migrate a part of the job by operating on individual containers, but these partial migrations did not occur in our experiments. The scheduler monitors hardware counters continuously, and so is able to detect any changes in the program behaviour.
This proof-of-concept scheduler is simple and does not take into account communication overhead that may occur if the processes of the same job run on several nodes. This is deferred to future research; in this work we show how to avoid contention-induced performance degradation.
The Balanced scheduler is implemented as a collection of daemons that run on each node. Within each node, jobs are scheduled by a user-level scheduler Clavis [18] that is based on the Distributed Intensity algorithm [5], [17], and assigns threads to cores so as to avoid multicore resource contention. Clavis is used as the intra-node scheduling policy under Bestfit and Min-collocation policies as well.
Our experimental environment mimics an HPC cluster. We do not have exclusive access to the actual cluster where we are able to modify the scheduling algorithm, so instead we used three identical multicore systems connected by the Gigabit Ethernet. We did not have multiple Intel systems, so our minicluster is comprised of the three AMD systems (Table I). Each system has two multicore CPUs with six cores each.
In order to be able to migrate MPI processes from one node to another after the job had begun execution, we place

Runtime relative to solo fds0 tachyon fds1 fds2 fds0 tachyon fds1 fds2 fds0 tachyon fds1 fds2

Best-fit Min-Collocation Balanced
Balanced (DI model) Balanced (our model)

Experiment 1

Node 1

Node 2

fds0, tachyon fds1, fds2

fds0, fds2

tachyon

fds0, tachyon fds1

Experiment 2

Node 1

Node 2

fds, tachyon zeus1

fds, tachyon zeus1, zeus2

TABLE IV JOB ALLOCATION ACROSS NODES

Node 3
fds1 fds2
Node 3 zeus2

200% 150% 100%
50% 0%

2.91 kWh

3.51 kWh

3.17 kWh

the processes into OpenVZ containers [19], which is a lightweight virtualization option for Linux. OpenVZ produced the lowest overhead compared to Xen and KVM, and offered better reliability than MPI checkpoints. The degradation for the job is estimated by averaging the degradation estimates for the corresponding containers.
We show two experiments demonstrating the benefits of the Balanced policy and the underlying model. The first experiment shows that the Balanced scheduler is able to respect performance degradation threshold, unlike the Bestfit scheduler, while using less energy than the Min-collocation scheduler. The second experiment shows that the Balanced scheduler that uses our model saves energy relative to the same scheduler that estimates performance degradation using a simple heuristic model proposed in the earlier work [5]. In both experiments we run four MPI jobs from the SPEC MPI suite, each with six processes. We report the running times and the energy consumed to run the workload4. The model was trained on SPEC CPU 2006 application; it did not include the MPI applications that we test.
Experiment 1: Improved performance fidelity. In this experiment, we run three copies of the fds application and one copy of tachyon. Fds is memory-intensive, so it would suffer performance degradation under contention, while tachyon would not. The Balanced scheduler is configured to avoid performance degradation above 50%. Table IV shows how the jobs were assigned to servers under the three algorithms, and Figure 6 shows the running time (relative to solo) and energy consumption. The red line indicates the 50% degradation threshold. Migration overhead is always included into the running times that we report.
Under Best-fit and Min-collocation, only one copy of fds is able to meet the 50% degradation constraint. The other two copies of fds suffer roughly 70% performance degradation, because they run together on the same server. Both schedulers assign jobs to nodes according to the order of their arrival. The jobs arrive in the order fds0, tachyon, fds1, fds2, so Bestfit fills the first node with fds0 and tachyon, then fills the second node with the other two copies of fds. Min-collocation assigns one job per node, but when fds2 arrives and all three nodes are filled, it is forced to scheduled fds2 with fds0 on Node 1. Although Min-collocation has an additional node at
4Energy was measured using the Dell Remote Access Control interface on our servers.

Running time (seconds)

Best-Fit

Min-Collocation

Balanced

Fig. 6. Performance and energy consumption during Experiment 1.

25000 20000

5.06 kWh

3.27 kWh

15000

10000

5000

0 fds tachyon zeus0 zeus1 fds tachyon zeus0 zeus1

Balanced -- DI Model

Balanced -- Our Model

Fig. 7. Performance and energy consumption during Experiment 2.

its disposal, it is unable to realize that it is better to run a copy of fds alone on Node 3 rather than tachyon.
Balanced, on the other hand, discovers that two copies of fds co-located on the same node will suffer more than 50% performance degradation and migrates one of them to the third node. As a result, it improves performance by about 15% on average (across all applications) relative to Best-fit and Min-collocation, while using 44% less energy than Mincollocation. Even though Balanced uses the same number of nodes as Min-collocation, it enables the workload to complete quicker, hence smaller energy consumption. Best-fit uses 8% less energy than Balanced, because it uses fewer nodes, but unlike Balanced it does not meet the 50% degradation threshold.
Experiment 2: Improved power efficiency. The purpose of this experiment is to demonstrate the benefit of precise estimates of performance degradation that would be produced by our model, as opposed to coarse estimates that would be produced by heuristic-based models [4], [5]. We compare the Balanced scheduler that uses our model with the same scheduler, but that uses the miss-rate based model underlying the Distributed Intensity (DI) algorithm. We refer to this version as Balanced-DI. The DI model checks if any colocated jobs have the miss rate greater than one miss per thousand instructions. If that threshold is exceeded, the jobs are deemed “contentious” and the scheduler distributes them to different CPUs, or in our case, different servers. The DI model is considered state-of-the-art, as most software-only contention aware algorithms relied on the models almost identical to DI [3], [4], [17].

We run the following jobs: fds, tachyon and two copies of zeus. Like in the first experiment, the schedulers are configured to avoid the degradation above 50%. Table IV shows how the schedulers assign the jobs to nodes. Balanced-DI observes that zeus has the miss rate of 25 misses per 1000 instructions, which by far exceeds its thresholds, so it migrates one copy of zeus to the third node. However, it turns out that despite its high miss rate, zeus experiences only negligible degradation when co-scheduled with another copy (see Figure 7). The Balanced scheduler that uses our model is able to produce an accurate estimate, so it does not migrate zeus to another node, and meets the degradation threshold while using 35% less energy than Balanced-DI.
V. RELATED WORK
Our work is the first to evaluate machine learning for modeling performance degradation on multicore CPUs. Besides machine learning, there are three major strategies that attacked the same problem: analytical modeling (often requiring unconventional hardware), models based on heuristics, and trialand-error methods.
Analytical modeling. One of the first models for resource contention on multicore chips was proposed by Chandra et al. [1]. It estimated the increase in the last-level cache (LLC) miss rate resulting from cache contention. Chandra’s model required unconventional hardware which in limited cases could be substituted with compiler extensions. The main limitation of this model is that it focused only on caches and did not address other resources, such as memory buses, system request queues, hardware prefetchers, etc., contention for which was found to be a crucial factor in performance degradation on modern CPUs [5], [17]. Machine learning models will capture contention in any hardware component as long as this component is represented by relevant performance events.
Eyerman, Hoste and Eeckhout [20] used a semi-manual methodology for modeling CPI stacks. They estimated unknown relationships using regression analysis. However, at the heart of their method is a generic analytical model for the processor. As we explained, we wanted to find a practical method that does not involve any manual model construction, and machine learning answered these needs.
Luque et al. developed a method to precisely count how many extra cycles the thread is wasting, waiting for CPU resources that are occupied as a result of contention [21]. This information can be used directly to estimate the performance degradation that contention is causing. While this is a very promising technique in terms of accuracy, it requires changing the hardware. Furthermore, this technique, at the time of this writing, addresses only shared caches. Our goal was to design a method that will work on today’s hardware and cover all kinds of shared CPU resources.
Models based on heuristics. In recent studies, the last-level cache miss rate was used as a heuristic to predict whether threads or processes sharing a multicore CPU are suffering performance degradation [3]–[5], [17]. In that work, the LLC miss rate was used to decide when the threads should be

scheduled on separate chips to avoid cache contention. While suitable for coarse-grained scheduling decisions, the miss rate is not sufficient to estimate performance degradation with a greater precision.
Furthermore, relying on a single indicator of performance (the miss-rate) to estimate the effect of sharing multiple resources is a fragile strategy. It may work as long as memory controllers and prefetch bandwidth are key contended resources on multicore systems [5], but if the hardware bottlenecks change, the heuristic will stop working. Furthermore, this method does not easily allow integration of other shared resources into the model. Machine learning can adjust to changes in hardware and be extended to model any new resources that emerge as important for contention; therefore, it is a more future-proof method.
Trial-and-error methods. Trial-and-error methods require running the workloads in various combinations (co-schedules) with other workloads [9] or with dummy benchmarks [7]. The goal is to observe how performance degradation changes in different co-schedules and to use that information to create online a machine- and workload-specific model of the degradation. A system called Cuanta is a very elegant solution, relying on a set of “clones”, each with a particular cache access pattern. By co-scheduling all clones with a target application, we can find the one that most closely mimics the behaviour of that application. Then, based on a previously constructed degradation matrix and application clones we can predict the degradation for any pair of applications. This approach works well when the number of cores per chip is small, but as the number grows, we would need to run a larger and larger combination of clones concurrently with the application. This is not practical, because the cores are unavailable to run other applications when we use them to run clones. Our machine learning model, on the other hand, requires hardware counter values that can be measured for all cores in parallel, and so the time to perform the on-line measurement does not grow with the number of cores.
VI. CONCLUSIONS
Our study aimed to investigate the effectiveness of machine learning in modeling contention-induced performance degradation: online, on a live workload, and without a priori knowledge of applications or the need to run them in isolation. We aimed for a model that seamlessly ports across different systems, and machine learning met this need as it does not rely on microarchitectural knowledge. We found that machine learning can indeed be used to built reasonably accurate models, which estimate degradation within 16% of the true value on average, however inaccurate estimates can occur if the test application is very different from the applications in the training set. Fortunately, these cases can be anticipated by checking how “dissimilar” the test application is from the training set, in terms of its attribute values. Our proposed method, the confidence predictor, successfully anticipates when the model is likely to produce an inaccurate estimate and reduces the maximum error by up to a factor of three.
ModelDegradationMachinePerformance DegradationApplications