APEX File Storage for Microsoft Azure: VM-Level Bursting Performance Impact
Fri, 05 Jul 2024 13:00:00 -0000
|Read Time: 0 minutes
In this blog, we’ll explore sequential read performance when leveraging virtual machine (VM)-level bursting.
For Azure VMs that support bursting and have VM-level bursting enabled by default, Microsoft Azure will start with fully stocked credits for the VM and allow bursting for up to 30 minutes at the maximum burst rate, which is higher than the virtual machine-level’s max uncached disk throughput, as shown in Table 1.
Table 1. Azure storage throughput limits and network bandwidth limits for tested node types
Node type/VM size | vCPU | Memory (GiB) | Max uncached disk throughput (MBps) | Max burst uncached disk throughput (MBps) | Max network bandwidth (Mbps) |
32 | 128 | 865 | 2,000 | 16,000 | |
48 | 192 | 1,315 | 3,000 | 24,000 | |
64 | 256 | 1,735 | 3,000 | 30,000 | |
96 | 384 | 2,600 | 4,000 | 35,000 | |
104 | 672 | 4,000 | 4,000 | 100,000 |
The VM-level burst credits are restocked whenever throughput falls below the VM-level maximum uncached disk throughput limit. It takes less than a day to fully restock when burst credits are fully depleted. For more information about virtual machine-level bursting, see Microsoft Azure Managed Disk Bursting.
Since sequential writes do not utilize virtual machine-level bursting due to the sequential write throughput being lower than the virtual machine-level’s max uncached disk throughput, virtual machine-level bursting does not affect sequential write performance.
Sequential read throughput
Figure 1 represents a 128KB sequential read workload with and without VM-level bursting:
- With VM-level bursting: The sequential read performance can surpass the VM-level maximum uncached disk throughput limit when utilizing VM-level bursting. With VM-level bursting, Figure 1 shows that a single D32ds_v5 node can exceed the VM-level maximum uncached disk throughput limit. However, the sequential read performance does not reach the VM-level maximum burst uncached disk throughput limit due to constraints imposed by the VM-level network bandwidth limit. This network bandwidth is shared between both the VM's external (front-end) and internal (back-end) interfaces.
- Without VM-level bursting: When VM-level burst credits are depleted, the sequential read workload runs without VM-level bursting. The sequential read throughput per node closely aligns with its VM-level maximum uncached disk throughput. Without VM-level bursting, Figure 1 shows that a single D32ds_v5 node read throughput is close to its VM-level maximum uncached disk throughput limit.
Note: This test uses a 4 D32ds_v5 nodes cluster with 12 data disks per node.
Author: Jason He, Principal Engineering Technologist