Grid GPUs - Xen Test

GPU-accelerated computing is the use of a graphics processing unit (GPU) together with a CPU to accelerate scientific, analytics, engineering, consumer, and enterprise applications. GPU-accelerated computing enhances application performance by offloading compute-intensive portions of the application to the GPU, while the remainder of the code still runs on the CPU. Architecturally, while a CPU has only few cores and handles few hundred threads at a time, a GPU is composed of hundreds of cores that can handle thousands of threads simultaneously and render a flawless rich graphics experience.

Now, imagine if you could access your GPU-accelerated applications, even those requiring intensive graphics power, anywhere on any device. NVIDIA GRID makes this possible. With NVIDIA GRID, a virtualized GPU designed specifically for virtualized server environments, data center managers can bring true PC graphics-rich experiences to users.

The NVIDIA GRID GPUs will be hosted in enterprise data centers and allow users to run virtual desktops or virtual applications on multiple devices connected to the internet and across multiple operating systems, including PCs, notebooks, tablets and even smartphones. Users can utilize their online-connected devices to enjoy the GPU power remotely.

In VDI/virtualized server environments, the NVIDIA GRID delivers GPU resources to virtual desktops/VMs. This way, graphics can be rendered on a virtual machine’s (VM’s) host server rather than on a physical end-point device. This technology now makes it possible to use virtual desktop technology to support users accessing graphics intensive workloads. There are two modes of making GPU resources available to virtual desktops:

  • Dedicated GPU or GPU Pass-through Technology: NVIDIA GPU pass-through technology lets you create a virtual workstation that gives users all the benefits of a dedicated graphics processor at their desk. By directly connecting a dedicated GPU to a virtual machine through the hypervisor, you can now allocate the full GPU and graphics memory capability to a single virtual machine without any resource compromise.

    Figure 1 : Dedicated GPU Technology

  • Shared GPU or Virtual GPU (vGPU) Technology: GRID vGPU is the industry's most advanced technology for sharing true GPU hardware acceleration between multiple virtual desktops—without compromising the graphics experience. With GRID vGPU technology, the graphics commands of each virtual machine are passed directly to the GPU, without translation by the hypervisor. This allows the GPU hardware to be time-sliced to deliver improved shared virtualized graphics performance. The GRID vGPU manager allows for management of user profiles. IT managers can assign the optimal amount of graphics memory and deliver a customized graphics profile to meet the specific needs of each user. Every virtual desktop has dedicated graphics memory, just like they would at their desk, so they always have the resources they need to launch and run their applications.

    Figure 2 : Shared vGPU Technology

    In GPU-enabled Citrix XenServer environments, if users to VMs/virtual desktops complain of slowness when accessing graphic applications, administrators must be able to instantly figure out what is causing the slowness – is it because adequate GPU resources are not available to the host? Or is it because of excessive utilization of GPU memory and processing resources by a few VMs/virtual desktops on the host?  Accurate answers to these questions can help administrators determine whether/not:

  • The host is sized with sufficient GPU resources;
  • The GPUs are configured with enough graphics and BAR1 memory;
  • The GPU technology in use – i.e., the GPU Pass-through technology or the Shared vGPU technology – is ideal for the graphics processing requirements of the environment; 

Measures to right-size the host and fine-tune its GPU configuration can be initiated based on the results of this analysis. This is exactly what the Grid GPUs - Xen test helps administrators achieve! 

This test supports GPU monitoring for NVIDIA K1 and K2 Grids installed on a Citrix XenServer. Using the test, administrators can monitor each physical GPU card installed on the server and can determine how actively memory on that card is utilized, thus revealing the card on which memory is used consistently. In addition, the test also indicates how busy each GPU card is, and in the process pinpoints those GPU cards that are being over-utilized by the VMs/virtual desktops on the host. The adequacy of the physical GPU resources is thus revealed. Moreover, detailed diagnostics provided by the test also lead you to those VMs/virtual desktops that are using each card. In addition, the power consumption and temperature of each GPU card is also monitored, so that its current temperature and power usage can be ascertained; administrators are thus alerted to abnormal power usage of the GPU and unexpected fluctuations in its temperature. The power limit set and the clock frequencies configured are also revealed, so that administrators can figure out whether the GPU is rightly configured for optimal processing or is any fine-tuning required.

Note:

This test will run and report GPU usage metrics only if the Shared GPU or Virtual GPU (vGPU) Technology is used to deliver GPU resources to virtual desktops/VMs. Where the Dedicated GPU or Pass-through GPU Technology is used, this test will report only the Virtual machines and the Mode metrics.

Target of the test : A XenServer host

Agent deploying the test : An internal/remote agent

Outputs of the test : One set of results for each GRID physical GPU assigned to the XenServer host being monitored. Additionally, measures are reported for the summary descriptor too. For the summary descriptor, this test will report the values aggregated across shared GPUs and those that are configured as Unclassified.

Configurable parameters for the test
  1. Test period - How often should the test be executed
  2. Host - The host for which the test is to be configured.
  3. PORT - The port at which the specified HOST listens. By default, this is NULL.
  4. XEN user - To enable the eG agent to connect to the XenServer API for collecting statistics of interest, this test should login to the XenServer as a root user. Provide the name of the root user in the XEN USER text box. Root user privileges are mandatory when monitoring a XenServer 5.5 (or below). However, if you are monitoring XenServer 5.6 (or above) and you prefer not to expose the credentials of the root user, then, you have the option of configuring a user with pool-admin privileges as the xen user. If you do not want to expose the credentials of a root/pool-admin user, then you can configure the tests with the credentials of a xen user with Read-only privileges to the XenServer. However, if this is done, then the Xen Uptime test will not run, and the Xen CPU and Xen Memory tests will not be able to report metrics for the control domain descriptor. To avoid such an outcome, do the following before attempting to configure the eG tests with a xen user who has Read-only privileges to the XenServer:

    • Modify the target XenServer’s configuration in the eG Enterprise system. For this, follow the Infrastructure -> Components -> Add/Modify menu sequence, pick Citrix XenServer as the Component type, and click the Modify button corresponding to the target XenServer.
    • In the modify component details page that then appears, make sure that the os is set to Xen and the Mode is set to ssh.
    • Then, in the same page, proceed to provide the User and Password of a user who has the right to connect to the XenServer console via SSH.
    • Then, click the Update button to save the changes.
  5. Once this is done, you can configure the eG tests with the credentials of a xen user with Read-only privileges.   

  6. xen password - The password of the specified xen user needs to be mentioned here.
  7. confirm password - Confirm the xen password by retyping it here.
  8. ssl - By default, the Xen Server is not SSL-enabled. This indicates that by default, the eG agent communicates with the XenServer using HTTP. Accordingly, the ssl flag is set to No by default. If you configure the XenServer to use SSL, then make sure that the SSL flag is set to Yes, so that the eG agent communicates with the XenServer using HTTPS. Note that a default SSL certificate comes bundled with every XenServer installation. If you want the eG agent to use this default certificate for communicating with an SSL-enabled XenServer, then no additional configuration is required. However, if you do not want to use the default certificate, then you can generate a self-signed certificate for use by the XenServer. In such a case, you need to explicitly follow the broad steps given below to enable the eG agent to communicate with the XenServer via HTTPS:

    • Obtain the server-certificate for the XenServer
    • Import the server-certificate into the local certificate store of the eG agent

    For a detailed discussion on each of these steps, refer to the Troubleshooting section of this document.

  9. webport - By default, in most virtualized environments, the XenServer listens on port 80 (if not SSL-enabled) or on port 443 (if SSL-enabled). This implies that while monitoring an SSL-enabled XenServer, the eG agent, by default, connects to port 443 of the server to pull out metrics, and while monitoring a non-SSL-enabled XenServer, the eG agent connects to port 80. Accordingly, the webport parameter is set to 80 or 443 depending upon the status of the ssl flag.  In some environments however, the default ports 80 or 443 might not apply. In such a case, against the webport parameter, you can specify the exact port at which the XenServer in your environment listens so that the eG agent communicates with that port.
Measurements made by the test
Measurement Description Measurement Unit Interpretation

GPU memory utilization:

Indicates the proportion of time over the past sample period during which global (device) memory was being read or written on this GPU.

Percent

A value close to 100% is a cause for concern as it indicates that graphics memory on a GPU is almost always in use.

In a Shared vGPU environment on the other hand, memory may be consumed all the time, if one/more VMs/virtual desktops on the host utilize the graphics memory excessively and constantly. If you find that only a single VM/virtual desktop has been consistently hogging the graphic memory resources, you may want to switch to the Dedicated GPU mode, so that excessive memory usage by that VM/virtual desktop has no impact on the performance of other VMs/virtual desktops on that host.

If the value of this measure is high almost all the time for most of the GPUs, it could mean that the host is not sized with adequate graphics memory.   

Used frame buffer memory:

Indicates the amount of frame buffer memory on-board this GPU that is being used by the host.

MiB

Frame buffer memory refers to the memory used to hold pixel properties such as color, alpha, depth, stencil, mask, etc.

Properties like the screen resolution, color level, and refresh speed of the frame buffer can impact graphics performance.

Also, if Error-correcting code (ECC) is enabled on a host, the available frame buffer memory may be decreased by several percent. This is because, ECC uses up memory to detect and correct the most common kinds of internal data corruption. Moreover, the driver may also reserve a small amount of memory for internal use, even without active work on the GPU; this too may impact frame buffer memory.

For optimal graphics performance therefore, adequate frame buffer memory should be allocated to the host. 

Free frame buffer memory:

Indicates the amount of frame buffer memory on-board this GPU that is yet to be used by the host.

MiB

 

Frame buffer memory utilization:

Indicates the percentage of frame buffer memory on-board this GPU that is being utilized by the host.

Percent

A value close to 100% is indicative of excessive frame buffer memory usage.

Properties like the screen resolution, color level, and refresh speed of the frame buffer can impact graphics performance.

Also, if Error-correcting code (ECC) is enabled on a host, the available frame buffer memory may be decreased by several percent. This is because, ECC uses up memory to detect and correct the most common kinds of internal data corruption. Moreover, the driver may also reserve a small amount of memory for internal use, even without active work on the GPU; this too may impact frame buffer memory.

For optimal graphics performance therefore, adequate frame buffer memory should be allocated to the host.

Use the detailed diagnosis of this measure to know the frame buffer memory usage of each VM on the monitored server. This will point you to those VMs that are draining the frame buffer memory.

GPU compute utilization:

Indicates the proportion of time over the past sample period during which one or more kernels was executing on this GPU.

Percent

A value close to 100% indicates that the GPU is busy processing graphic requests almost all the time.

In a Shared vGPU environment on the other hand, a GPU may be in use almost all the time, if one/more VMs/virtual desktops on the host are running highly graphic-intensive applications. If you find that only a single VM/virtual desktop has been consistently hogging the GPU resources, you may want to switch to the Dedicated GPU mode, so that excessive GPU usage by that VM/virtual desktop has no impact on the performance of other VMs/virtual desktops on that host.

If all GPUs are found to be busy most of the time, you may want to consider augmenting the GPU resources of the host. 

Compare the value of this measure across physical GPUs to know which GPU is being used more than the rest. 

Power consumption:

Indicates the current power usage of this GPU.

Watts

A very high value is indicative of excessive power usage by the GPU.

In such cases, you may want to enable Power management so that the GPU limits power draw under load to fit within a predefined power envelope by manipulating the current performance state.

Core GPU temperature:

Indicates the current temperature of this GPU.

Celsius

Ideally, the value of this measure should be low. A very high value is indicative of abnormal GPU temperature.

Total framebuffer memory:

Indicates the total size of frame buffer memory of this GPU.

 

MB

Frame buffer memory refers to the memory used to hold pixel properties such as color, alpha, depth, stencil, mask, etc.

Total BAR1 memory:

Indicates the total size of the BAR1 memory of this GPU.

MiB

BAR1 is used to map the frame buffer (device memory) so that it can be directly accessed by the CPU or by 3rd party devices (peer-to-peer on the PCIe bus).

Used BAR1 memory:

Indicates the amount of BAR1 memory on this GPU that is currently being used by the host.

MiB

For better user experience with graphic applications, enough BAR1 memory should be available to the host.

Free BAR1 memory:

Indicates the amount of BAR1 memory of this GPU that is still to be used by the host.

MiB

 

BAR1 memory utilization:

Indicates the percentage of the total BAR1 memory on this GPU that is currently being utilized by the host.

Percent

A value close to 100% is indicative of excessive BAR1 memory usage by the host.

For best graphics performance, sufficient BAR1 memory resources should be available to the host.

Power management:

Indicates whether/not power management is enabled for this GPU.

 

Many NVIDIA graphics cards support multiple performance levels so that the server can save power when full graphics performance is not required. 

The default Power Management Mode of the graphics card is Adaptive. In this mode, the graphics card monitors GPU usage and seamlessly switches between modes based on the performance demands of the application. This allows the GPU to always use the minimum amount of power required to run a given application. This mode is recommended by NVIDIA for best overall balance of power and performance. If the power management mode is set to Adaptive, the value of this measure will be Supported.

Alternatively, you can set the Power Management Mode to Maximum Performance. This mode allows users to maintain the card at its maximum performance level when 3D applications are running regardless of GPU usage. If the power management mode of a GPU is Maximum Performance, then the value of this measure will be Maximum.

The numeric values that correspond to these measure values are discussed in the table below:

Measure Value Numeric Value

Supported

1

Maximum

0

Note:

By default, this measure will report the Measure Values listed in the table above to indicate the power management status. In the graph of this measure however, the same is represented using the numeric equivalents only.

 

Power limit:

Indicates the power limit configured for this GPU.

Watts

This measure will report a value only if the value of the ‘Power management’ measure is ‘Supported’.

The power limit setting controls how much voltage a GPU can use when under load. Its not advisable to set the power limit at its maximum – i.e., the value of this measure should not be the same as the value of the Max power limit measure - as it can cause the GPU to behave strangely under duress.

Default power limit:

Indicates the default power management algorithm's power ceiling for this GPU.

Watts

This measure will report a value only if the value of the ‘Power management’ measure is ‘Supported’.

 

Enforced power limit:

Indicates the power management algorithm's power ceiling for this GPU.

Watts

This measure will report a value only if the value of the ‘Power management’ measure is ‘Supported’.

The total board power draw is manipulated by the power management algorithm such that it stays under the value reported by this measure.

Min power limit:

The minimum value that the power limit of this GPU can be set to.

Watts

This measure will report a value only if the value of the ‘Power management’ measure is ‘Supported’.

 

Max power limit:

The maximum value that the power limit of this GPU can be set to.

Watts

This measure will report a value only if the value of the ‘Power management’ measure is ‘Supported’.

If the value of this measure is the same as that of the Power limit measure, then the GPU may behave strangely.

Graphics clock:

Indicates the current frequency of the graphics clock of this GPU.

MHz

GPU has many more cores than your average CPU but these cores are much simpler and much smaller so that many more actually fit on a small piece of silicon. These smaller, simpler cores go by different names depending upon the tasks they perform. Stream processors are the cores that perform a single thread at a slow rate. But since GPUs contain numerous stream processors, they make overall computation high.

The streaming multiprocessor clock refers to how fast the stream processors run. The Graphics clock is the speed at which the GPU operates. The memory clock is how fast the memory on the card runs.

By correlating the frequencies of these clocks (i.e., the value of these measures) with the memory usage, power usage, and overall performance of the GPU, you can figure out if overclocking is required or not. 

Overclocking is the process of forcing a GPU core/memory to run faster than its manufactured frequency. Overclocking can have both positive and negative effects on GPU performance. For instance, memory overclocking helps on cards with low memory bandwidth, and with games with a lot of post-processing/textures/filters like AA that are VRAM intensive. On the other hand, speeding up the operation frequency of a shader/streaming processor/memory, without properly analyzing its need and its effects, may increase its thermal output in a linear fashion. At the same time, boosting voltages will cause the generated heat to sky rocket. If improperly managed, these increases in temperature can cause permanent physical damage to the core/memory or even “heat death”.

Putting an adequate cooling system into place, adjusting the power provided to the GPU, monitoring your results with the right tools and doing the necessary research are all critical steps on the path to safe and successful overclocking.  

Streaming multiprocessor clock:

Indicates the current frequency of the streaming multiprocessor clock of this GPU.

MHz

Memory clock:

Indicates the current frequency of the memory clock of this GPU.

MHz

Fan speed:

Indicates the percent of maximum speed that this GPU’s fan is currently intended to run at.

Percent

The value of this measure could range from 0 to 100%.

An abnormally high value for this measure could indicate a problem condition – eg., a sudden surge in the temperature of the GPU that could cause the fan to spin faster.

Note that the reported speed is only the intended fan speed. If the fan is physically blocked and unable to spin, this output will not match the actual fan speed. Many parts do not report fan speeds because they rely on cooling via fans in the surrounding enclosure. By default the fan speed is increased or decreased automatically in response to changes in temperature.

Compute processes:

Indicates the number of processes having compute context on this GPU.

Number

Use the detailed diagnosis of this measure to know which processes are currently using the GPU. The process details provided as part of the detailed diagnosis include, the PID of the process, the process name, and the GPU memory used by the process.

Note that the GPU memory usage of the processes will not be available in the detailed diagnosis, if the Windows platform on which XenApp operates is running in the WDDM mode. In this mode, the Windows KMD manages all the memory, and not the NVIDIA driver. Therefore, the NVIDIA SMI commands that the test uses to collect metrics will not be able to capture the GPU memory usage of the processes.  

Volatile single bit errors:

Indicates the number of volatile single bit errors in this GPU.

Number

Volatile error counters track the number of errors detected since the last driver load. Single bit ECC errors are automatically corrected by the hardware and do not result in data corruption.

Ideally, the value of this measure should be 0.

Volatile double bit errors:

Indicates the total number of volatile double bit errors in this GPU.

Number

Volatile error counters track the number of errors detected since the last driver load. Double bit errors are detected but not corrected.

Ideally, the value of this measure should be 0.

Aggregate single bit errors:

Indicates the total number of aggregate single bit errors in this GPU.

Number

Aggregate error counts persist indefinitely and thus act as a lifetime counter. Single bit ECC errors are automatically corrected by the hardware and do not result in data corruption.

Ideally, the value of this measure should be 0.

Aggregate double bit errors:

Indicates the total number of aggregate double bit errors in this GPU.

Number

Aggregate error counts persist indefinitely and thus act as a lifetime counter. Double bit errors are detected but not corrected.

Ideally, the value of this measure should be 0.

Virtual machines:

Indicates the number of virtual machines allocated to this GPU Grid card.

Number

If the Mode measure reports the value Pass-through, then the value of this measure will be 1. Using the detailed diagnosis of this measure, you can identify the VMs that are using the GPU grid card and the process utilization of each VM.

Mode:

Indicates the mode using which the GPU resources were delivered to the VMs.

 

The values that this measure can take and their corresponding numeric values are as follows:

Measure Value Numeric Values

Pass through

0

Shared

1

Unavailable (GPU card is not allocated to any VM)

2

Note:

By default, this test reports the Measure Values listed in the table above to indicate the mode of GPU delivery. In the graph of this measure however, the same is represented using the numeric equivalents only.

Remaining capacity:

Indicates the number of vGPUs in this GPU grid card that are yet to be allocated to the VMs on this host. 

Number

Ideally, the value of this measure should be high.

A value close to 0 could indicate that vGPUs have been over-allocated to the VMs. This could result in a serious GPU-contention on the host.