Resource Allocation and Performance on Anypoint Runtime Fabric

You are responsible for ensuring that you remain compliant with your MuleSoft licensing. To remain compliant, ensure one of the following:

For instructions on how to track Reserved vCPU and vCPU Limit, refer to How to Get Runtime Fabric (RTF) CPU Reserved and Limit for Self Managed Kubernetes and RTF Appliance.

For instructions on how to monitor your usage and get alerts if your usage exceeds your license, refer to Cronjob Script to Monitor RTF Appliance CPU of Mule Apps from the MuleSoft Knowledge Base.

When the Reserved vCPU and vCPU Limit are equal, the CPU on the worker node is allocated in a guaranteed model. The container always has the specified amount of CPU available. The vCPU is reserved for each replica of the application and is guaranteed regardless of what else is running on Runtime Fabric.

Guaranteed CPU allocation allows applications to perform consistently by ensuring there is no CPU contention between other applications. It guarantees performance in all cases.

When the value of vCPU Limit is set higher than the value of Reserved vCPU, the application can burst up to the either the vCPU limit value or the amount of unreserved vCPU on the worker node, whichever is less.

When estimating the CPU and memory required for worker nodes, Anypoint Monitoring agent requires at least 0.05 vCPU cores and 50 Mi of memory for Anypoint Monitoring. This container is used to stream application metrics to the control plane and cannot be disabled. These core and memory allocations are not used for Mule licensing.

When using CPU bursting, you must consider the following factors:

Resource intensive applications might require additional CPU capacity. If an application deployment fails using the minimum vCPU limit, increase the vCPU limit and redeploy the application. * The CPU limit is upper-bounded by the CPU cores provided on the worker nodes. * The maximum recommended utilization per CPU core is 20 - 25 simple applications and API gateways. More complex applications with higher usage may need more capacity. * Runtime Fabric on VMs / Bare Metal runs a small number of services on worker nodes. These services consume between 0.3 and 0.5 vCPU, depending on whether log forwarding is enabled for Anypoint Monitoring. * Applications compete for the unallocated CPU remaining on their worker node. As more applications are deployed on a worker node, less unallocated CPU is available to share with other applications and can result in poor performance. The following strategies can be considered for minimizing CPU contention with your applications:

The minimum values for the amount of memory allocated to each replica of a Mule application or API gateway are the following:

When Anypoint Monitoring is enabled, 50MB memory is reserved for running the Anypoint Monitoring sidecar containers, which can burst up to 70MB for each application deployed when processing metrics data.

Anypoint Platform allocates native and heap memory for a deployed application. Heap memory is the portion of the total memory that is made available to the Mule runtime and the application. Heap memory is used for tasks like handling payload.

Both CloudHub and Anypoint Runtime Fabric allocate both types of memory. However, they differ in how the memory allocation of each memory type is calculated.

The startup time for a Mule applications is correlated to the total number of vCPU cores the application has access to:

The resources allocated to your Mule application determine the application’s performance. Following are the approximate values for throughput based on the total number of vCPU cores allocated for a single Mule application performing simple processing on a 10-KB payload:

Internal load balancer memory requirements are impacted by the number of threads, response time latency, and message sizes. Use the following guidelines when allocating memory:

Inbound traffic is processed using an internal load balancer managed by Anypoint Runtime Fabric on VMs / Bare Metal. Because this load balancer is responsible for TLS termination, the number of resources required scales based on the number of incoming connections and the average payload size for each request.

Performance test results are based on an Runtime Fabric on VMs / Bare Metal cluster with controller nodes using AWS EC2 M4 instances. The cluster was configured with three m4.large controller nodes and three r4.large worker nodes. The load generator used in the performance test was hosted on a separate AWS instance in the same region. The M4 instances featured a custom Intel Xeon E5-2676 v3 Haswell processor optimized specifically for EC2, which ran at a base clock rate of 2.4 GHz. Using Intel Turbo Boost increased the clock rate to go as high as 3.0 GHz.

A load generator based on C++, which has better efficiency in SSL connections was used to yield the maximum throughput.

The following table summarizes the approximate requests (averaging 10 KB) that can be served with a single replica of the internal load balancer, based on the number of CPU cores. In most cases, Elliptical Curve Digital Signature Algorithm (ECDSA) provides double the performance of a 2K RSA key. Supported curves are secp521r1 (P-521), secp384r1 (P-384), and secp256r1, also known as prime256v1 (P-256).

*Doubling the RSA key length degrades performance by at least a factor of 6.

Ensure that you allocate enough CPU resources to support a minmum of 10 PEM/P12 or 8 JKS/JCEKS certificates. The number of recommended cores are:

RSA keys are the most common type of keys. RSA keys of 2K length offer the best compromise between security and performance.

ECDSA keys are also supported. In most cases, ECDSA doubles the performance of a 2K RSA key. Supported curves are: