Detecting Memory Bandwidth Saturation in Threaded Applications

Detecting Memory Bandwidth Saturation in Threaded Applications (PDF 231KB)


Memory sub-system components contribute significantly to the performance characteristics of an application. As an increasing number of threads or processes share the limited resources of cache capacity and memory bandwidth, the scalability of a threaded application can become constrained. Memory-intensive threaded applications can suffer from memory bandwidth saturation as more threads are introduced. In such cases, the threaded application won’t scale as expected, and performance can be reduced. This article introduces techniques to detect memory bandwidth saturation in threaded applications.

This article is part of the larger series, "Intel Guide for Developing Multithreaded Applications," which provides guidelines for developing efficient multithreaded applications for Intel® platforms.


As modern processors include more cores and bigger caches, they become faster at a higher rate than the memory sub-system components. The increasing core count on a per-die basis has put pressure on the cache capacity and memory bandwidth. As a result, optimally utilizing the available cache and memory bandwidth to each core is essential in developing forward-scaling applications. If a system isn't capable of moving data from main memory to the cores fast enough, the cores will sit idle as they wait for the data to arrive. An idle core during computation is a wasted resource that increases the overall execution time of the computation and will negate some of the benefits of more cores.

The current generation of Intel® processors based on the Nehalem architecture moved from the traditional front-side-bus (FSB) approach to non-uniform memory access/architecture (NUMA) model to increase the available memory bandwidth to the cores and reduce the bandwidth saturation issues mentioned above. Figure 1 depicts the FSB to NUMA transition.

Figure 1. Transition from FSB to NUMA.

The clear symptom of bandwidth saturation for any parallel application is non-scaling behavior. In other words, an application that has saturated the available memory bandwidth will not scale effectively to more threads or cores. However, there are many causes for multi-threaded applications not to scale and some of these performance inhibiting factors include threading overhead, synchronization overhead, load imbalance, and inappropriate granularity. Intel® Thread Profiler is designed to identify such performance issues at the application level.

The following results are taken after running the STREAM benchmark version 5.6 with various numbers of threads (only triad scores are shown).



Rate (MB/s)

Avg time

Min time

Max time

1 Thread






2 Threads






4 Threads






It is easy to see that STREAM does not benefit from having more threads on this particular platform (a single-socket Intel® Core™ 2 Quad-based system). Closer inspection of the results shows that even though there was a slight increase in the triad score for the two-thread version, the four-thread version performed even worse than the single threaded run.

Figure 2 shows Intel Thread Profiler analysis of the benchmark. The timeline view reveals that all threads are perfectly balanced and have no synchronization overheads. While it is a powerful tool for identifying threading performance issues at application level, Intel Thread Profiler will not detect memory bandwidth saturation in threaded applications.

Figure 2. Intel Thread Profiler timeline view of the STREAM benchmark using four OpenMP* threads.


Intel® VTune™ Performance Analyzer and Performance Tuning Utility (PTU) used in combination with event-based sampling (EBS), can help developers measure application bandwidth usage, which can then be checked against the achievable (or theoretical) bandwidth on the system. Event-based sampling relies on the performance monitoring unit (PMU) supported by the processors.

VTune analyzer and PTU can help developers estimate the memory bandwidth usage of a particular application by using EBS. On Intel® Core™ 2 microarchitecture CPU_CLK_UNHALTED.CORE and BUS_TRANS_MEM.ALL_AGENTS performance events can be used to estimate the memory bandwidth.

  • The CPU_CLK_UNHALTED.CORE event counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction.
  • The BUS_TRANS_MEM.ALL_AGENTS event counts activity initiated by any agent on the bus. In systems where each processor is attached to a different bus, the count reflects only the activity for the bus on which the processor resides.

On Core 2-based systems memory bandwidth can be estimated by using the following formula:


Figure 3. VTune analyzer EBS analysis of STREAM with four threads.

Figure 3 shows the EBS results of the STREAM benchmark when four threads were used. By using the above formula, it is possible to estimate the memory bandwidth usage of STREAM as 7.6Gb/sec.

Memory Bandwidth = (64 * 1,419,200,000 * 2.9GHz) / 35,576,000,000 = 7.6GB/sec

The STREAM-reported sustainable Triad score was 7.7GB/seconds, so the VTune analyzer-based calculation is quite reasonable. The STREAM benchmark was chosen to demonstrate how memory bandwidth measured using EBS can approximately measure the achievable memory bandwidth on a particular system.

If an application doesn’t scale when more threads are added to take advantage of the available cores, and if Intel Thread Profiler doesn’t show any application-level threading problems as mentioned above, then the following three steps can help the user determine whether or not a particular application is saturating the available memory bandwidth:

  1. Run STREAM or similar benchmarks to get an idea of the sustainable memory bandwidth on the target system.
  2. Run the target application under VTune analyzer or PTU and collect the appropriate performance counters using EBS. For Core 2 microarchitecture, these events are again CPU_CLK_UNHALTED.CORE and BUS_TRANS_MEM.ALL_AGENTS (Formula 1).
  3. Compare VTune analyzer-measured memory bandwidth numbers to the sustainable or achievable memory bandwidth measured in step 1. If the application is saturating the available bandwidth, then this particular application won’t scale with more cores.

Generally speaking, a memory-bound application (one whose performance is limited by the memory access speed) won’t benefit from having more threads.

Usage Guidelines

The new Intel® Core™ i7 and Xeon® 5500 series processors are referred to as having an "uncore, " which is that part of the processor that is external to all the individual cores. For example, the Intel Core i7 processor has four cores that share an L3 cache and a memory interface. The L3 and memory interface are considered to be part of the uncore (see Figure 4).

Neither the VTune analyzer nor PTU support the sampling of events that are triggered in the uncore of the processor, and the memory bandwidth measurement must be performed differently. The relevant performance events used for measuring bandwidth are not sampled using EBS as is usual with VTune analyzer or PTU; rather, they are counted using time-based sampling. This means that the bandwidth is measured for the entire system over a designated time range, and it isn’t possible to see how much of the bandwidth usage comes from specific functions, processes, and modules.

Figure 4. Simplified block diagram of a 4 core Nehalem processor.

The formula given above can be used to measure the memory bandwidth usage of any application, module, or function on Core 2 based systems except on Core 2 based Xeon MP processors, which also have uncore parts. The basic formula for measuring the memory bandwidth on Nehalem architecture-based systems can be given as follows:

Memory Bandwidth = 1.0e-9 * (UNC_IMC_NORMAL_READS.ANY+UNC_IMC_WRITES.FULL.ANY)*64 / (wall clock time in seconds)

Additional Resources

For more complete information about compiler optimizations, see our Optimization Notice.


Sidharth Kashyap's picture

Circa 2016

The  BUS_TRANS_MEM.ALL_AGENTS event is not listed in VTune and Uncore events are available. 

One of the methods to get the Memory Bandwidth (that worked for me):


Manuel Selva's picture

Hi all,

Can someone explain the link with Event Based Sampling (EBS) when memory bandwidth is etimated with VTune analyzer and PTU on the core 2 micro architecture. I only see there two events count and no event sampling. Is event sampling used to count only events coming from the considered STREAM application ? How is this done ?


Roman Dementiev (Intel)'s picture

regarding the vmware ESX comment: VMware Fusion 5 supports now Intel VTune Amplifier event sampling:

anonymous's picture

Is there some good formula by which we can estimate memory bandwidth on Sandy Bridge microachetecture ?


anonymous's picture

Are there options available for detecting front-side bus saturation on Intel servers runnign VMware ESX operating systems? Seems like a lacking area for performance monitoring with VMware. Have to wait till systems begin to drag and then just guess if that is the case.



anonymous's picture

Note that the STREAM benchmark prints the best result for each kernel across the multiple iterations run (default=10, with the first iteration always excluded), while the performance counters in the example above appear to be based on sampling over the entire benchmark execution. Depending on the particular mode of operation, performance counts may or may not be linked back to source code sections, exposing the counts to additional biases including program data initialization, execution of all iterations of all four kernels, and validation of the results. This can easily account for the small (<5%) difference between the 4-thread TRIAD result of 7780 MB/s and the performance counter result of 7404 MB/s. Computing the bandwidth using the weighted average execution time would probably give a much closer match, and inline access to the performance counters (measuring each kernel independently) can give extremely accurate and repeatable results.

Caution should be applied when interpreting uncore results from short program sections. Write-back caches send the data to memory when the cache overflows, not when the data is written. In some cases the eventual writes of the data to DRAM can be missed if the performance monitors are stopped too soon. This caution applies to all systems with write-back caches (which is pretty much all systems from all vendors). Just to make things complicated, this issue does not apply to highly optimized versions of the STREAM benchmark on some architectures (including those discussed in this article), since these can use "streaming stores" to send the writes directly to DRAM rather than into the cache hierarchy. You can usually identify the presence of "streaming store" optimizations in the STREAM benchmark when there is no significant difference between the average of the Copy & Scale kernels and the average of the Add & Triad kernels. Without "streaming store" optimizations, systems with write-back caches typically have to read the target array before writing it, resulting in a 50% increase in memory operations for the Copy and Scale kernels and a 33% increase in memory operations for the Add and Triad kernels.

Add a Comment

Have a technical question? Visit our forums. Have site or software product issues? Contact support.