| title | description | ms.topic | ms.date | ms.custom |
|---|---|---|---|---|
Automatically scale compute nodes in an Azure Batch pool |
Enable automatic scaling on a cloud pool to dynamically adjust the number of compute nodes in the pool. |
how-to |
04/13/2021 |
H1Hack27Feb2017, fasttrack-edit, devx-track-csharp |
Azure Batch can automatically scale pools based on parameters that you define, saving you time and money. With automatic scaling, Batch dynamically adds nodes to a pool as task demands increase, and removes compute nodes as task demands decrease.
To enable automatic scaling on a pool of compute nodes, you associate the pool with an autoscale formula that you define. The Batch service uses the autoscale formula to determine how many nodes are needed to execute your workload. These nodes may be dedicated nodes or low-priority nodes. Batch will then periodically review service metrics data and use it to adjust the number of nodes in the pool based on your formula and at an interval that you define.
You can enable automatic scaling when you create a pool, or apply it to an existing pool. Batch enables you to evaluate your formulas before assigning them to pools and to monitor the status of automatic scaling runs. Once you configure a pool with automatic scaling, you can make changes to the formula later.
Important
When you create a Batch account, you can specify the pool allocation mode, which determines whether pools are allocated in a Batch service subscription (the default) or in your user subscription. If you created your Batch account with the default Batch service configuration, then your account is limited to a maximum number of cores that can be used for processing. The Batch service scales compute nodes only up to that core limit. For this reason, the Batch service may not reach the target number of compute nodes specified by an autoscale formula. See Quotas and limits for the Azure Batch service for information on viewing and increasing your account quotas.
If you created your account with user subscription mode, then your account shares in the core quota for the subscription. For more information, see Virtual Machines limits in Azure subscription and service limits, quotas, and constraints.
An autoscale formula is a string value that you define that contains one or more statements. The autoscale formula is assigned to a pool's autoScaleFormula element (Batch REST) or CloudPool.AutoScaleFormula property (Batch .NET). The Batch service uses your formula to determine the target number of compute nodes in the pool for the next interval of processing. The formula string can't exceed 8 KB, can include up to 100 statements that are separated by semicolons, and can include line breaks and comments.
You can think of automatic scaling formulas as a Batch autoscale "language." Formula statements are free-formed expressions that can include both service-defined variables (defined by the Batch service) and user-defined variables. Formulas can perform various operations on these values by using built-in types, operators, and functions. For example, a statement might take the following form:
$myNewVariable = function($ServiceDefinedVariable, $myCustomVariable);
Formulas generally contain multiple statements that perform operations on values that are obtained in previous statements. For example, first we obtain a value for variable1, then pass it to a function to populate variable2:
$variable1 = function1($ServiceDefinedVariable);
$variable2 = function2($OtherServiceDefinedVariable, $variable1);
Include these statements in your autoscale formula to arrive at a target number of compute nodes. Dedicated nodes and low-priority nodes each have their own target settings. An autoscale formula can include a target value for dedicated nodes, a target value for low-priority nodes, or both.
The target number of nodes may be higher, lower, or the same as the current number of nodes of that type in the pool. Batch evaluates a pool's autoscale formula at a specific automatic scaling intervals. Batch adjusts the target number of each type of node in the pool to the number that your autoscale formula specifies at the time of evaluation.
Below are examples of two autoscale formulas, which can be adjusted to work for most scenarios. The variables startingNumberOfVMs and maxNumberofVMs in the example formulas can be adjusted to your needs.
With this autoscale formula, the pool is initially created with a single VM. The $PendingTasks metric defines the number of tasks that are running or queued. The formula finds the average number of pending tasks in the last 180 seconds and sets the $TargetDedicatedNodes variable accordingly. The formula ensures that the target number of dedicated nodes never exceeds 25 VMs. As new tasks are submitted, the pool automatically grows. As tasks complete, VMs become free and the autoscaling formula shrinks the pool.
This formula scales dedicated nodes, but can be modified to apply to scale low-priority nodes as well.
startingNumberOfVMs = 1;
maxNumberofVMs = 25;
pendingTaskSamplePercent = $PendingTasks.GetSamplePercent(180 * TimeInterval_Second);
pendingTaskSamples = pendingTaskSamplePercent < 70 ? startingNumberOfVMs : avg($PendingTasks.GetSample(180 * TimeInterval_Second));
$TargetDedicatedNodes=min(maxNumberofVMs, pendingTaskSamples);
$NodeDeallocationOption = taskcompletion;
This example creates a pool that starts with 25 low-priority nodes. Every time a low-priority node is preempted, it is replaced with a dedicated node. As with the first example, the maxNumberofVMs variable prevents the pool from exceeding 25 VMs. This example is useful for taking advantage of low-priority VMs while also ensuring that only a fixed number of preemptions will occur for the lifetime of the pool.
maxNumberofVMs = 25;
$TargetDedicatedNodes = min(maxNumberofVMs, $PreemptedNodeCount.GetSample(180 * TimeInterval_Second));
$TargetLowPriorityNodes = min(maxNumberofVMs , maxNumberofVMs - $TargetDedicatedNodes);
$NodeDeallocationOption = taskcompletion;
You'll learn more about how to create autoscale formulas and see additional example autoscale formulas later in this topic.
You can use both service-defined and user-defined variables in your autoscale formulas.
The service-defined variables are built in to the Batch service. Some service-defined variables are read-write, and some are read-only.
User-defined variables are variables that you define. In the example formula shown above, $TargetDedicatedNodes and $PendingTasks are service-defined variables, while startingNumberOfVMs and maxNumberofVMs are user-defined variables.
Note
Service-defined variables are always preceded by a dollar sign ($). For user-defined variables, the dollar sign is optional.
The following tables show the read-write and read-only variables that are defined by the Batch service.
You can get and set the values of these service-defined variables to manage the number of compute nodes in a pool.
| Variable | Description |
|---|---|
| $TargetDedicatedNodes | The target number of dedicated compute nodes for the pool. This is specified as a target because a pool may not always achieve the desired number of nodes. For example, if the target number of dedicated nodes is modified by an autoscale evaluation before the pool has reached the initial target, the pool may not reach the target. A pool in an account created in Batch service mode may not achieve its target if the target exceeds a Batch account node or core quota. A pool in an account created in user subscription mode may not achieve its target if the target exceeds the shared core quota for the subscription. |
| $TargetLowPriorityNodes | The target number of low-priority compute nodes for the pool. This specified as a target because a pool may not always achieve the desired number of nodes. For example, if the target number of low-priority nodes is modified by an autoscale evaluation before the pool has reached the initial target, the pool may not reach the target. A pool may also not achieve its target if the target exceeds a Batch account node or core quota. For more information on low-priority compute nodes, see Use low-priority VMs with Batch. |
| $NodeDeallocationOption | The action that occurs when compute nodes are removed from a pool. Possible values are:
|
Note
The $TargetDedicatedNodes variable can also be specified using the alias $TargetDedicated. Similarly, the $TargetLowPriorityNodes variable can be specified using the alias $TargetLowPriority. If both the fully named variable and its alias are set by the formula, the value assigned to the fully named variable will take precedence.
You can get the value of these service-defined variables to make adjustments that are based on metrics from the Batch service.
Important
Job release tasks are not currently included in variables that provide task counts, such as $ActiveTasks and $PendingTasks. Depending on your autoscale formula, this can result in nodes being removed with no nodes available to run job release tasks.
| Variable | Description |
|---|---|
| $CPUPercent | The average percentage of CPU usage. |
| $WallClockSeconds | The number of seconds consumed. |
| $MemoryBytes | The average number of megabytes used. |
| $DiskBytes | The average number of gigabytes used on the local disks. |
| $DiskReadBytes | The number of bytes read. |
| $DiskWriteBytes | The number of bytes written. |
| $DiskReadOps | The count of read disk operations performed. |
| $DiskWriteOps | The count of write disk operations performed. |
| $NetworkInBytes | The number of inbound bytes. |
| $NetworkOutBytes | The number of outbound bytes. |
| $SampleNodeCount | The count of compute nodes. |
| $ActiveTasks | The number of tasks that are ready to execute but are not yet executing. This includes all tasks that are in the active state and whose dependencies have been satisfied. Any tasks that are in the active state but whose dependencies have not been satisfied are excluded from the $ActiveTasks count. For a multi-instance task, $ActiveTasks will include the number of instances set on the task. |
| $RunningTasks | The number of tasks in a running state. |
| $PendingTasks | The sum of $ActiveTasks and $RunningTasks. |
| $SucceededTasks | The number of tasks that finished successfully. |
| $FailedTasks | The number of tasks that failed. |
| $TaskSlotsPerNode | The number of task slots that can be used to run concurrent tasks on a single compute node in the pool. |
| $CurrentDedicatedNodes | The current number of dedicated compute nodes. |
| $CurrentLowPriorityNodes | The current number of low-priority compute nodes, including any nodes that have been preempted. |
| $PreemptedNodeCount | The number of nodes in the pool that are in a preempted state. |
Tip
These read-only service-defined variables are objects that provide various methods to access data associated with each. For more information, see Obtain sample data later in this article.
Note
Use $RunningTasks when scaling based on the number of tasks running at a point in time, and $ActiveTasks when scaling based on the number of tasks that are queued up to run.
Autoscale formulas support the following types:
- double
- doubleVec
- doubleVecList
- string
- timestamp--a compound structure that contains the following members:
- year
- month (1-12)
- day (1-31)
- weekday (in the format of number; for example, 1 for Monday)
- hour (in 24-hour number format; for example, 13 means 1 PM)
- minute (00-59)
- second (00-59)
- timeinterval
- TimeInterval_Zero
- TimeInterval_100ns
- TimeInterval_Microsecond
- TimeInterval_Millisecond
- TimeInterval_Second
- TimeInterval_Minute
- TimeInterval_Hour
- TimeInterval_Day
- TimeInterval_Week
- TimeInterval_Year
These operations are allowed on the types that are listed in the previous section.
| Operation | Supported operators | Result type |
|---|---|---|
| double operator double | +, -, *, / | double |
| double operator timeinterval | * | timeinterval |
| doubleVec operator double | +, -, *, / | doubleVec |
| doubleVec operator doubleVec | +, -, *, / | doubleVec |
| timeinterval operator double | *, / | timeinterval |
| timeinterval operator timeinterval | +, - | timeinterval |
| timeinterval operator timestamp | + | timestamp |
| timestamp operator timeinterval | + | timestamp |
| timestamp operator timestamp | - | timeinterval |
| operator double | -, ! | double |
| operator timeinterval | - | timeinterval |
| double operator double | <, <=, ==, >=, >, != | double |
| string operator string | <, <=, ==, >=, >, != | double |
| timestamp operator timestamp | <, <=, ==, >=, >, != | double |
| timeinterval operator timeinterval | <, <=, ==, >=, >, != | double |
| double operator double | &&, || | double |
When testing a double with a ternary operator (double ? statement1 : statement2), nonzero is true, and zero is false.
You can use these predefined functions when defining an autoscale formula.
| Function | Return type | Description |
|---|---|---|
| avg(doubleVecList) | double | Returns the average value for all values in the doubleVecList. |
| len(doubleVecList) | double | Returns the length of the vector that is created from the doubleVecList. |
| lg(double) | double | Returns the log base 2 of the double. |
| lg(doubleVecList) | doubleVec | Returns the component-wise log base 2 of the doubleVecList. A vec(double) must be explicitly passed for the parameter. Otherwise, the double lg(double) version is assumed. |
| ln(double) | double | Returns the natural log of the double. |
| ln(doubleVecList) | doubleVec | Returns the natural log of the double. |
| log(double) | double | Returns the log base 10 of the double. |
| log(doubleVecList) | doubleVec | Returns the component-wise log base 10 of the doubleVecList. A vec(double) must be explicitly passed for the single double parameter. Otherwise, the double log(double) version is assumed. |
| max(doubleVecList) | double | Returns the maximum value in the doubleVecList. |
| min(doubleVecList) | double | Returns the minimum value in the doubleVecList. |
| norm(doubleVecList) | double | Returns the two-norm of the vector that is created from the doubleVecList. |
| percentile(doubleVec v, double p) | double | Returns the percentile element of the vector v. |
| rand() | double | Returns a random value between 0.0 and 1.0. |
| range(doubleVecList) | double | Returns the difference between the min and max values in the doubleVecList. |
| std(doubleVecList) | double | Returns the sample standard deviation of the values in the doubleVecList. |
| stop() | Stops evaluation of the autoscaling expression. | |
| sum(doubleVecList) | double | Returns the sum of all the components of the doubleVecList. |
| time(string dateTime="") | timestamp | Returns the time stamp of the current time if no parameters are passed, or the time stamp of the dateTime string if that is passed. Supported dateTime formats are W3C-DTF and RFC 1123. |
| val(doubleVec v, double i) | double | Returns the value of the element that is at location i in vector v, with a starting index of zero. |
Some of the functions that are described in the previous table can accept a list as an argument. The comma-separated list is any combination of double and doubleVec. For example:
doubleVecList := ( (double | doubleVec)+(, (double | doubleVec) )* )?
The doubleVecList value is converted to a single doubleVec before evaluation. For example, if v = [1,2,3], then calling avg(v) is equivalent to calling avg(1,2,3). Calling avg(v, 7) is equivalent to calling avg(1,2,3,7).
You can use both resource and task metrics when you're defining a formula. You adjust the target number of dedicated nodes in the pool based on the metrics data that you obtain and evaluate. For more information on each metric, see the Variables section above.
| Metric | Description |
|---|---|
| Resource | Resource metrics are based on the CPU, the bandwidth, the memory usage of compute nodes, and the number of nodes. These service-defined variables are useful for making adjustments based on node count:
These service-defined variables are useful for making adjustments based on node resource usage:
|
| Task | Task metrics are based on the status of tasks, such as Active, Pending, and Completed. The following service-defined variables are useful for making pool-size adjustments based on task metrics:
|
The core operation of an autoscale formula is to obtain task and resource metric data (samples), and then adjust pool size based on that data. As such, it's important to have a clear understanding of how autoscale formulas interact with samples.
Autoscale formulas act on samples of metric data provided by the Batch service. A formula will grow or shrink the pool size based on the values that it obtains. Service-defined variables are objects that provide methods to access data that is associated with that object. For example, the following expression shows a request to get the last five minutes of CPU usage:
$CPUPercent.GetSample(TimeInterval_Minute * 5)
The following methods may be used to obtain sample data about service-defined variables.
| Method | Description |
|---|---|
| GetSample() | The GetSample() method returns a vector of data samples.A sample is 30 seconds worth of metrics data. In other words, samples are obtained every 30 seconds. But as noted below, there is a delay between when a sample is collected and when it is available to a formula. As such, not all samples for a given time period may be available for evaluation by a formula.
|
| GetSamplePeriod() | Returns the period of samples that were taken in a historical sample data set. |
| Count() | Returns the total number of samples in the metric history. |
| HistoryBeginTime() | Returns the time stamp of the oldest available data sample for the metric. |
| GetSamplePercent() | Returns the percentage of samples that are available for a given time interval. For example, doubleVec GetSamplePercent( (timestamp or timeinterval) startTime [, (timestamp or timeinterval) endTime] ). Because the GetSample method fails if the percentage of samples returned is less than the samplePercent specified, you can use the GetSamplePercent method to check first. Then you can perform an alternate action if insufficient samples are present, without halting the automatic scaling evaluation. |
The Batch service periodically takes samples of task and resource metrics and makes them available to your autoscale formulas. These samples are recorded every 30 seconds by the Batch service. However, there is typically a delay between when those samples were recorded and when they are made available to (and can be read by) your autoscale formulas. Additionally, samples may not be recorded for a particular interval because of factors such as network or other infrastructure issues.
When samplePercent is passed to the GetSample() method or the GetSamplePercent() method is called, percent refers to a comparison between the total possible number of samples that are recorded by the Batch service and the number of samples that are available to your autoscale formula.
Let's look at a 10-minute timespan as an example. Because samples are recorded every 30 seconds within that 10-minute timespan, the maximum total number of samples recorded by Batch would be 20 samples (2 per minute). However, due to the inherent latency of the reporting mechanism and other issues within Azure, there may be only 15 samples that are available to your autoscale formula for reading. So, for example, for that 10-minute period, only 75% of the total number of samples recorded may be available to your formula.
Your autoscale formulas will grow and shrink your pools by adding or removing nodes. Because nodes cost you money, be sure that your formulas use an intelligent method of analysis that is based on sufficient data. We recommend that you use a trending-type analysis in your formulas. This type grows and shrinks your pools based on a range of collected samples.
To do so, use GetSample(interval look-back start, interval look-back end) to return a vector of samples:
$runningTasksSample = $RunningTasks.GetSample(1 * TimeInterval_Minute, 6 * TimeInterval_Minute);
When the above line is evaluated by Batch, it returns a range of samples as a vector of values. For example:
$runningTasksSample=[1,1,1,1,1,1,1,1,1,1];
Once you've collected the vector of samples, you can then use functions like min(), max(), and avg() to derive meaningful values from the collected range.
For additional security, you can force a formula evaluation to fail if less than a certain sample percentage is available for a particular time period. When you force a formula evaluation to fail, you instruct Batch to cease further evaluation of the formula if the specified percentage of samples is not available. In this case, no change is made to the pool size. To specify a required percentage of samples for the evaluation to succeed, specify it as the third parameter to GetSample(). Here, a requirement of 75 percent of samples is specified:
$runningTasksSample = $RunningTasks.GetSample(60 * TimeInterval_Second, 120 * TimeInterval_Second, 75);
Because there may be a delay in sample availability, you should always specify a time range with a look-back start time that is older than one minute. It takes approximately one minute for samples to propagate through the system, so samples in the range (0 * TimeInterval_Second, 60 * TimeInterval_Second) may not be available. Again, you can use the percentage parameter of GetSample() to force a particular sample percentage requirement.
Important
We strongly recommend that you avoid relying only on GetSample(1) in your autoscale formulas. This is because GetSample(1) essentially says to the Batch service, "Give me the last sample you have, no matter how long ago you retrieved it." Since it is only a single sample, and it may be an older sample, it may not be representative of the larger picture of recent task or resource state. If you do use GetSample(1), make sure that it's part of a larger statement and not the only data point that your formula relies on.
You build an autoscale formula by forming statements that use the above components, then combine those statements into a complete formula. In this section, we create an example autoscale formula that can perform real-world scaling decisions and make adjustments.
First, let's define the requirements for our new autoscale formula. The formula should:
- Increase the target number of dedicated compute nodes in a pool if CPU usage is high.
- Decrease the target number of dedicated compute nodes in a pool when CPU usage is low.
- Always restrict the maximum number of dedicated nodes to 400.
- When reducing the number of nodes, don't remove nodes that are running tasks; if necessary, wait until tasks have finished before removing nodes.
The first statement in our formula will increase the number of nodes during high CPU usage. We'll define a statement that populates a user-defined variable ($totalDedicatedNodes) with a value that is 110 percent of the current target number of dedicated nodes, but only if the minimum average CPU usage during the last 10 minutes was above 70 percent. Otherwise, it uses the value for the current number of dedicated nodes.
$totalDedicatedNodes =
(min($CPUPercent.GetSample(TimeInterval_Minute * 10)) > 0.7) ?
($CurrentDedicatedNodes * 1.1) : $CurrentDedicatedNodes;
To decrease the number of dedicated nodes during low CPU usage, the next statement in our formula sets the same $totalDedicatedNodes variable to 90 percent of the current target number of dedicated nodes, if average CPU usage in the past 60 minutes was under 20 percent. Otherwise, it uses the current value of $totalDedicatedNodes that we populated in the statement above.
$totalDedicatedNodes =
(avg($CPUPercent.GetSample(TimeInterval_Minute * 60)) < 0.2) ?
($CurrentDedicatedNodes * 0.9) : $totalDedicatedNodes;
Now, we'll limit the target number of dedicated compute nodes to a maximum of 400.
$TargetDedicatedNodes = min(400, $totalDedicatedNodes)
Finally, we'll ensure that nodes aren't removed until their tasks are finished.
$NodeDeallocationOption = taskcompletion;
Here's the complete formula:
$totalDedicatedNodes =
(min($CPUPercent.GetSample(TimeInterval_Minute * 10)) > 0.7) ?
($CurrentDedicatedNodes * 1.1) : $CurrentDedicatedNodes;
$totalDedicatedNodes =
(avg($CPUPercent.GetSample(TimeInterval_Minute * 60)) < 0.2) ?
($CurrentDedicatedNodes * 0.9) : $totalDedicatedNodes;
$TargetDedicatedNodes = min(400, $totalDedicatedNodes)
$NodeDeallocationOption = taskcompletion;
Note
If you choose to, you can include both comments and line breaks in formula strings. Also be aware that missing semicolons may result in evaluation errors.
By default, the Batch service adjusts a pool's size according to its autoscale formula every 15 minutes. This interval is configurable by using the following pool properties:
- CloudPool.AutoScaleEvaluationInterval (Batch .NET)
- autoScaleEvaluationInterval (REST API)
The minimum interval is five minutes, and the maximum is 168 hours. If an interval outside this range is specified, the Batch service returns a Bad Request (400) error.
Note
Autoscaling is not currently intended to respond to changes in less than a minute, but rather is intended to adjust the size of your pool gradually as you run a workload.
Pool autoscaling can be configured using any of the Batch SDKs, the Batch REST API Batch PowerShell cmdlets, and the Batch CLI. In this section, you can see examples for both .NET and Python.
To create a pool with autoscaling enabled in .NET, follow these steps:
- Create the pool with BatchClient.PoolOperations.CreatePool.
- Set the CloudPool.AutoScaleEnabled property to
true. - Set the CloudPool.AutoScaleFormula property with your autoscale formula.
- (Optional) Set the CloudPool.AutoScaleEvaluationInterval property (default is 15 minutes).
- Commit the pool with CloudPool.Commit or CommitAsync.
The following example creates an autoscale-enabled pool in .NET. The pool's autoscale formula sets the target number of dedicated nodes to 5 on Mondays, and to 1 on every other day of the week. The automatic scaling interval is set to 30 minutes. In this and the other C# snippets in this article, myBatchClient is a properly initialized instance of the BatchClient class.
CloudPool pool = myBatchClient.PoolOperations.CreatePool(
poolId: "mypool",
virtualMachineSize: "standard_d1_v2",
VirtualMachineConfiguration: new VirtualMachineConfiguration(
imageReference: new ImageReference(
publisher: "MicrosoftWindowsServer",
offer: "WindowsServer",
sku: "2019-datacenter-core",
version: "latest"),
nodeAgentSkuId: "batch.node.windows amd64");
pool.AutoScaleEnabled = true;
pool.AutoScaleFormula = "$TargetDedicatedNodes = (time().weekday == 1 ? 5:1);";
pool.AutoScaleEvaluationInterval = TimeSpan.FromMinutes(30);
await pool.CommitAsync();Important
When you create an autoscale-enabled pool, don't specify the targetDedicatedNodes parameter or the targetLowPriorityNodes parameter on the call to CreatePool. Instead, specify the AutoScaleEnabled and AutoScaleFormula properties on the pool. The values for these properties determine the target number of each type of node.
To manually resize an autoscale-enabled pool (for example, with BatchClient.PoolOperations.ResizePoolAsync), you must first disable automatic scaling on the pool, then resize it.
To create autoscale-enabled pool with the Python SDK:
- Create a pool and specify its configuration.
- Add the pool to the service client.
- Enable autoscale on the pool with a formula you write.
The following example illustrates these steps.
# Create a pool; specify configuration
new_pool = batch.models.PoolAddParameter(
id="autoscale-enabled-pool",
virtual_machine_configuration=batchmodels.VirtualMachineConfiguration(
image_reference=batchmodels.ImageReference(
publisher="Canonical",
offer="UbuntuServer",
sku="18.04-LTS",
version="latest"
),
node_agent_sku_id="batch.node.ubuntu 18.04"),
vm_size="STANDARD_D1_v2",
target_dedicated_nodes=0,
target_low_priority_nodes=0
)
batch_service_client.pool.add(new_pool) # Add the pool to the service client
formula = """$curTime = time();
$workHours = $curTime.hour >= 8 && $curTime.hour < 18;
$isWeekday = $curTime.weekday >= 1 && $curTime.weekday <= 5;
$isWorkingWeekdayHour = $workHours && $isWeekday;
$TargetDedicated = $isWorkingWeekdayHour ? 20:10;""";
# Enable autoscale; specify the formula
response = batch_service_client.pool.enable_auto_scale(pool_id, auto_scale_formula=formula,
auto_scale_evaluation_interval=datetime.timedelta(minutes=10),
pool_enable_auto_scale_options=None,
custom_headers=None, raw=False)Tip
More examples of using the Python SDK can be found in the Batch Python Quickstart repository on GitHub.
Each Batch SDK provides a way to enable automatic scaling. For example:
- BatchClient.PoolOperations.EnableAutoScaleAsync (Batch .NET)
- Enable automatic scaling on a pool (REST API)
When you enable autoscaling on an existing pool, keep in mind:
- If autoscaling is currently disabled on the pool, you must specify a valid autoscale formula when you issue the request. You can optionally specify an automatic scaling interval. If you don't specify an interval, the default value of 15 minutes is used.
- If autoscaling is currently enabled on the pool, you can specify a new formula, a new interval, or both. You must specify at least one of these properties.
- If you specify a new automatic scaling interval, the existing schedule is stopped and a new schedule is started. The new schedule's start time is the time at which the request to enable autoscaling was issued.
- If you omit either the autoscale formula or interval, the Batch service will continue to use the current value of that setting.
Note
If you specified values for the targetDedicatedNodes or targetLowPriorityNodes parameters of the CreatePool method when you created the pool in .NET, or for the comparable parameters in another language, then those values are ignored when the autoscale formula is evaluated.
This C# example uses the Batch .NET library to enable autoscaling on an existing pool.
// Define the autoscaling formula. This formula sets the target number of nodes
// to 5 on Mondays, and 1 on every other day of the week
string myAutoScaleFormula = "$TargetDedicatedNodes = (time().weekday == 1 ? 5:1);";
// Set the autoscale formula on the existing pool
await myBatchClient.PoolOperations.EnableAutoScaleAsync(
"myexistingpool",
autoscaleFormula: myAutoScaleFormula);To update the formula on an existing autoscale-enabled pool, call the operation to enable autoscaling again with the new formula. For example, if autoscaling is already enabled on myexistingpool when the following .NET code is executed, its autoscale formula is replaced with the contents of myNewFormula.
await myBatchClient.PoolOperations.EnableAutoScaleAsync(
"myexistingpool",
autoscaleFormula: myNewFormula);To update the autoscale evaluation interval of an existing autoscale-enabled pool, call the operation to enable autoscaling again with the new interval. For example, to set the autoscale evaluation interval to 60 minutes for a pool that's already autoscale-enabled in .NET:
await myBatchClient.PoolOperations.EnableAutoScaleAsync(
"myexistingpool",
autoscaleEvaluationInterval: TimeSpan.FromMinutes(60));You can evaluate a formula before applying it to a pool. This lets you test the formula's results before you put it into production.
Before you can evaluate an autoscale formula, you must first enable autoscaling on the pool with a valid formula, such as the one-line formula $TargetDedicatedNodes = 0. Then, use one of the following to evaluate the formula you want to test:
-
BatchClient.PoolOperations.EvaluateAutoScale or EvaluateAutoScaleAsync
These Batch .NET methods require the ID of an existing pool and a string containing the autoscale formula to evaluate.
-
Evaluate an automatic scaling formula
In this REST API request, specify the pool ID in the URI, and the autoscale formula in the autoScaleFormula element of the request body. The response of the operation contains any error information that might be related to the formula.
This Batch .NET example evaluates an autoscale formula. If the pool doesn't already use autoscaling, we enable it first.
// First obtain a reference to an existing pool
CloudPool pool = await batchClient.PoolOperations.GetPoolAsync("myExistingPool");
// If autoscaling isn't already enabled on the pool, enable it.
// You can't evaluate an autoscale formula on a non-autoscale-enabled pool.
if (pool.AutoScaleEnabled == false)
{
// We need a valid autoscale formula to enable autoscaling on the
// pool. This formula is valid, but won't resize the pool:
await pool.EnableAutoScaleAsync(
autoscaleFormula: "$TargetDedicatedNodes = $CurrentDedicatedNodes;",
autoscaleEvaluationInterval: TimeSpan.FromMinutes(5));
// Batch limits EnableAutoScaleAsync calls to once every 30 seconds.
// Because we want to apply our new autoscale formula below if it
// evaluates successfully, and we *just* enabled autoscaling on
// this pool, we pause here to ensure we pass that threshold.
Thread.Sleep(TimeSpan.FromSeconds(31));
// Refresh the properties of the pool so that we've got the
// latest value for AutoScaleEnabled
await pool.RefreshAsync();
}
// We must ensure that autoscaling is enabled on the pool prior to
// evaluating a formula
if (pool.AutoScaleEnabled == true)
{
// The formula to evaluate - adjusts target number of nodes based on
// day of week and time of day
string myFormula = @"
$curTime = time();
$workHours = $curTime.hour >= 8 && $curTime.hour < 18;
$isWeekday = $curTime.weekday >= 1 && $curTime.weekday <= 5;
$isWorkingWeekdayHour = $workHours && $isWeekday;
$TargetDedicatedNodes = $isWorkingWeekdayHour ? 20:10;
";
// Perform the autoscale formula evaluation. Note that this code does not
// actually apply the formula to the pool.
AutoScaleRun eval =
await batchClient.PoolOperations.EvaluateAutoScaleAsync(pool.Id, myFormula);
if (eval.Error == null)
{
// Evaluation success - print the results of the AutoScaleRun.
// This will display the values of each variable as evaluated by the
// autoscale formula.
Console.WriteLine("AutoScaleRun.Results: " +
eval.Results.Replace("$", "\n $"));
// Apply the formula to the pool since it evaluated successfully
await batchClient.PoolOperations.EnableAutoScaleAsync(pool.Id, myFormula);
}
else
{
// Evaluation failed, output the message associated with the error
Console.WriteLine("AutoScaleRun.Error.Message: " +
eval.Error.Message);
}
}Successful evaluation of the formula shown in this code snippet produces results similar to:
AutoScaleRun.Results:
$TargetDedicatedNodes=10;
$NodeDeallocationOption=requeue;
$curTime=2016-10-13T19:18:47.805Z;
$isWeekday=1;
$isWorkingWeekdayHour=0;
$workHours=0
To ensure that your formula is performing as expected, we recommend that you periodically check the results of the autoscaling runs that Batch performs on your pool. To do so, get (or refresh) a reference to the pool, then examine the properties of its last autoscale run.
In Batch .NET, the CloudPool.AutoScaleRun property has several properties that provide information about the latest automatic scaling run performed on the pool:
In the REST API, the Get information about a pool request returns information about the pool, which includes the latest automatic scaling run information in the autoScaleRun property.
The following C# example uses the Batch .NET library to print information about the last autoscaling run on pool myPool.
await Cloud pool = myBatchClient.PoolOperations.GetPoolAsync("myPool");
Console.WriteLine("Last execution: " + pool.AutoScaleRun.Timestamp);
Console.WriteLine("Result:" + pool.AutoScaleRun.Results.Replace("$", "\n $"));
Console.WriteLine("Error: " + pool.AutoScaleRun.Error);Sample output from the preceding example:
Last execution: 10/14/2016 18:36:43
Result:
$TargetDedicatedNodes=10;
$NodeDeallocationOption=requeue;
$curTime=2016-10-14T18:36:43.282Z;
$isWeekday=1;
$isWorkingWeekdayHour=0;
$workHours=0
Error:
You can also check automatic scaling history by querying PoolAutoScaleEvent. This event is emitted by Batch Service to record each occurrence of autoscale formula evaluation and execution, which can be helpful to troubleshoot potential issues.
Sample event for PoolAutoScaleEvent:
{
"id": "poolId",
"timestamp": "2020-09-21T23:41:36.750Z",
"formula": "...",
"results": "$TargetDedicatedNodes=10;$NodeDeallocationOption=requeue;$curTime=2016-10-14T18:36:43.282Z;$isWeekday=1;$isWorkingWeekdayHour=0;$workHours=0",
"error": {
"code": "",
"message": "",
"values": []
}
}Let's look at a few formulas that show different ways to adjust the amount of compute resources in a pool.
Suppose you want to adjust the pool size based on the day of the week and time of day. This example shows how to increase or decrease the number of nodes in the pool accordingly.
The formula first obtains the current time. If it's a weekday (1-5) and within working hours (8 AM to 6 PM), the target pool size is set to 20 nodes. Otherwise, it's set to 10 nodes.
$curTime = time();
$workHours = $curTime.hour >= 8 && $curTime.hour < 18;
$isWeekday = $curTime.weekday >= 1 && $curTime.weekday <= 5;
$isWorkingWeekdayHour = $workHours && $isWeekday;
$TargetDedicatedNodes = $isWorkingWeekdayHour ? 20:10;
$NodeDeallocationOption = taskcompletion;
$curTime can be adjusted to reflect your local time zone by adding time() to the product of TimeZoneInterval_Hour and your UTC offset. For instance, use $curTime = time() + (-6 * TimeInterval_Hour); for Mountain Daylight Time (MDT). Keep in mind that the offset would need to be adjusted at the start and end of daylight saving time (if applicable).
In this C# example, the pool size is adjusted based on the number of tasks in the queue. We've included both comments and line breaks in the formula strings.
// Get pending tasks for the past 15 minutes.
$samples = $PendingTasks.GetSamplePercent(TimeInterval_Minute * 15);
// If we have fewer than 70 percent data points, we use the last sample point,
// otherwise we use the maximum of last sample point and the history average.
$tasks = $samples < 70 ? max(0,$PendingTasks.GetSample(1)) : max( $PendingTasks.GetSample(1), avg($PendingTasks.GetSample(TimeInterval_Minute * 15)));
// If number of pending tasks is not 0, set targetVM to pending tasks, otherwise
// half of current dedicated.
$targetVMs = $tasks > 0? $tasks:max(0, $TargetDedicatedNodes/2);
// The pool size is capped at 20, if target VM value is more than that, set it
// to 20. This value should be adjusted according to your use case.
$TargetDedicatedNodes = max(0, min($targetVMs, 20));
// Set node deallocation mode - let running tasks finish before removing a node
$NodeDeallocationOption = taskcompletion;This C# example adjusts the pool size based on the number of tasks. This formula also takes into account the TaskSlotsPerNode value that has been set for the pool. This approach is useful in situations where parallel task execution has been enabled on your pool.
// Determine whether 70 percent of the samples have been recorded in the past
// 15 minutes; if not, use last sample
$samples = $ActiveTasks.GetSamplePercent(TimeInterval_Minute * 15);
$tasks = $samples < 70 ? max(0,$ActiveTasks.GetSample(1)) : max( $ActiveTasks.GetSample(1),avg($ActiveTasks.GetSample(TimeInterval_Minute * 15)));
// Set the number of nodes to add to one-fourth the number of active tasks
// (the TaskSlotsPerNode property on this pool is set to 4, adjust
// this number for your use case)
$cores = $TargetDedicatedNodes * 4;
$extraVMs = (($tasks - $cores) + 3) / 4;
$targetVMs = ($TargetDedicatedNodes + $extraVMs);
// Attempt to grow the number of compute nodes to match the number of active
// tasks, with a maximum of 3
$TargetDedicatedNodes = max(0,min($targetVMs,3));
// Keep the nodes active until the tasks finish
$NodeDeallocationOption = taskcompletion;This example shows a C# example with an autoscale formula that sets the pool size to a specified number of nodes for an initial time period. After that, it adjusts the pool size based on the number of running and active tasks.
Specifically, this formula does the following:
- Sets the initial pool size to four nodes.
- Does not adjust the pool size within the first 10 minutes of the pool's lifecycle.
- After 10 minutes, obtains the max value of the number of running and active tasks within the past 60 minutes.
- If both values are 0 (indicating that no tasks were running or active in the last 60 minutes), the pool size is set to 0.
- If either value is greater than zero, no change is made.
string now = DateTime.UtcNow.ToString("r");
string formula = string.Format(@"
$TargetDedicatedNodes = {1};
lifespan = time() - time(""{0}"");
span = TimeInterval_Minute * 60;
startup = TimeInterval_Minute * 10;
ratio = 50;
$TargetDedicatedNodes = (lifespan > startup ? (max($RunningTasks.GetSample(span, ratio), $ActiveTasks.GetSample(span, ratio)) == 0 ? 0 : $TargetDedicatedNodes) : {1});
", now, 4);- Learn how to execute multiple tasks simultaneously on the compute nodes in your pool. Along with autoscaling, this can help to lower job duration for some workloads, saving you money.
- Learn how to query the Azure Batch service efficiently for further efficiency.