संपादित करें

इसके माध्यम से साझा किया गया


Azure NetApp Files regular volume performance benchmarks for Linux

This article describes performance benchmarks Azure NetApp Files delivers for Linux with a regular volume.

Whole file streaming workloads (scale-out benchmark tests)

The intent of a scale-out test is to show the performance of an Azure NetApp File volume when scaling out (or increasing) the number of clients generating simultaneous workload to the same volume. These tests are generally able to push a volume to the edge of its performance limits and are indicative of workloads such as media rendering, AI/ML, and other workloads that utilize large compute farms to perform work.

High I/OP scale-out benchmark configuration

These benchmarks used the following:

  • A single Azure NetApp Files 100-TiB regular volume with a 1-TiB dataset using the Ultra performance tier
  • FIO (with and without setting randrepeat=0)
  • 4-KiB and 8-KiB block sizes
  • 6 D32s_v5 virtual machines running RHEL 9.3
  • NFSv3
  • Manual QoS
  • Mount options: rw,nconnect=8,hard,rsize=262144,wsize=262144,vers=3,tcp,bg

High throughput scale-out benchmark configuration

These benchmarks used the following:

  • A single Azure NetApp Files regular volume with a 1-TiB dataset using the Ultra performance tier FIO (with and without setting randrepeat=0)
  • FIO (with and without setting randrepeat=0)
  • 64-KiB and 256-KiB block size
  • 6 D32s_v5 virtual machines running RHEL 9.3
  • NFSv3
  • Manual QoS
  • Mount options: rw,nconnect=8,hard,rsize=262144,wsize=262144,vers=3,tcp,bg

Parallel network connection (nconnect) benchmark configuration

These benchmarks used the following:

  • A single Azure NetApp Files regular volume with a 1-TiB dataset using the Ultra performance tier
  • FIO (with and without setting randrepeat=0)
  • 4-KiB and 64-KiB wsize/rsize
  • A single D32s_v4 virtual machine running RHEL 9.3
  • NFSv3 with and without nconnect
  • Mount options: rw,nconnect=8,hard,rsize=262144,wsize=262144,vers=3,tcp,bg

Scale-up benchmark tests

The scale-up test’s intent is to show the performance of an Azure NetApp File volume when scaling up (or increasing) the number of jobs generating simultaneous workload across multiple TCP connections on a single client to the same volume (such as with nconnect).

Without nconnect, these workloads can't push the limits of a volume’s maximum performance, since the client can't generate enough IO or network throughput. These tests are generally indicative of what a single user’s experience might be in workloads such as media rendering, databases, AI/ML, and general file shares.

High I/OP scale-out benchmarks

The following benchmarks show the performance achieved for Azure NetApp Files with a high I/OP workload using:

  • 32 clients
  • 4-KiB and 8-KiB random reads and writes
  • 1-TiB dataset
  • Read/write ratios as follows: 100%:0%, 90%:10%, 80%:20%, and so on
  • With and without filesystem caching involved (using randrepeat=0 in FIO)

For more information, see Testing methodology.

Results: 4 KiB, random, client caching included

In this benchmark, FIO ran without the randrepeat option to randomize data. Thus, an indeterminate amount of caching came into play. This configuration results in slightly better overall performance numbers than tests run without caching with the entire IO stack being utilized.

In the following graph, testing shows an Azure NetApp Files regular volume can handle between approximately 130,000 pure random 4-KiB writes and approximately 460,000 pure random 4 KiB reads during this benchmark. Read-write mix for the workload adjusted by 10% for each run.

As the read-write I/OP mix increases towards write-heavy, the total I/OPS decrease.

Diagram of benchmark tests with 4 KiB, random, client caching included.

Results: 4 KiB, random, client caching excluded

In this benchmark, FIO was run with the setting randrepeat=0 to randomize data, reducing the caching influence on performance. This resulted in an approximately 8% reduction in write I/OPS and an approximately 17% reduction in read I/OPS, but displays performance numbers more representative of what the storage can actually do.

In the following graph, testing shows an Azure NetApp Files regular volume can handle between approximately 120,000 pure random 4-KiB writes and approximately 388,000 pure random 4-KiB reads. Read-write mix for the workload adjusted by 25% for each run.

As the read-write I/OP mix increases towards write-heavy, the total I/OPS decrease.

Diagram of benchmark tests with 4 KiB, random, client caching excluded.

Results: 8 KiB, random, client caching excluded

Larger read and write sizes will result in fewer total I/OPS, as more data can be sent with each operation. An 8-KiB read and write size was used to more accurately simulate what most modern applications use. For instance, many EDA applications utilize 8-KiB reads and writes.

In this benchmark, FIO ran with randrepeat=0 to randomize data so the client caching impact was reduced. In the following graph, testing shows that an Azure NetApp Files regular volume can handle between approximately 111,000 pure random 8-KiB writes and approximately 293,000 pure random 8-KiB reads. Read-write mix for the workload adjusted by 25% for each run.

As the read-write I/OP mix increases towards write-heavy, the total I/OPS decrease.

Diagram of benchmark tests with 8 KiB, random, client caching excluded.

Side-by-side comparisons

To illustrate how caching can influence the performance benchmark tests, the following graph shows total I/OPS for 4-KiB tests with and without caching mechanisms in place. As shown, caching provides a slight performance boost for I/OPS fairly consistent trending.

Diagram comparing 4 KiB benchmark tests.

Specific offset, streaming random read/write workloads: scale-up tests using parallel network connections (nconnect)

The following tests show a high I/OP benchmark using a single client with 4-KiB random workloads and a 1-TiB dataset. The workload mix generated uses a different I/O depth each time. To boost the performance for a single client workload, the nconnect mount option was used to improve parallelism in comparison to client mounts without the nconnect mount option.

When using a standard TCP connection that provides only a single path to the storage, fewer total operations are sent per second than when a mount is able to leverage more TCP connections (such as with nconnect) per mount point. When using nconnect, the total latency for the operations is generally lower. These tests are also run with randrepeat=0 to intentionally avoid caching. For more information on this option, see Testing methodology.

Results: 4 KiB, random, with and without nconnect, caching excluded

The following graphs show a side-by-side comparison of 4-KiB reads and writes with and without nconnect to highlight the performance improvements seen when using nconnect: higher overall I/OPS, lower latency.

Diagram of 4-KiB read performance.

Diagram of 4-KiB write performance.

High throughput benchmarks

The following benchmarks show the performance achieved for Azure NetApp Files with a high throughput workload.

High throughput workloads are more sequential in nature and often are read/write heavy with low metadata. Throughput is generally more important than I/OPS. These workloads typically leverage larger read/write sizes (64K to 256K), which generate higher latencies than smaller read/write sizes, since larger payloads will naturally take longer to be processed.

Examples of high throughput workloads include:

  • Media repositories
  • High performance compute
  • AI/ML/LLP

The following tests show a high throughput benchmark using both 64-KiB and 256-KiB sequential workloads and a 1-TiB dataset. The workload mix generated decreases a set percentage at a time and demonstrates what you can expect when using varying read/write ratios (for instance, 100%:0%, 90%:10%, 80%:20%, and so on).

Results: 64 KiB sequential I/O, caching included

In this benchmark, FIO ran using looping logic that more aggressively populated the cache, so an indeterminate amount of caching influenced the results. This results in slightly better overall performance numbers than tests run without caching.

In the graph below, testing shows that an Azure NetApp Files regular volume can handle between approximately 4,500MiB/s pure sequential 64-KiB reads and approximately 1,600MiB/s pure sequential 64-KiB writes. The read-write mix for the workload was adjusted by 10% for each run.

Diagram of 64-KiB benchmark tests with sequential I/O and caching included.

Results: 64 KiB sequential I/O, caching excluded

In this benchmark, FIO ran using looping logic that less aggressively populated the cache. Client caching didn't influence the results. This configuration results in slightly better write performance numbers, but lower read numbers than tests without caching.

In the following graph, testing demonstrates that an Azure NetApp Files regular volume can handle between approximately 3,600MiB/s pure sequential 64-KiB reads and approximately 2,400MiB/s pure sequential 64-KiB writes. During the tests, a 50/50 mix showed total throughput on par with a pure sequential read workload.

The read-write mix for the workload was adjusted by 25% for each run.

Diagram of 64-KiB benchmark tests with sequential I/O, caching excluded.

Results: 256 KiB sequential I/O, caching excluded

In this benchmark, FIO ran using looping logic that less aggressively populated the cache, so caching didn't influence the results. This configuration results in slightly less write performance numbers than 64-KiB tests, but higher read numbers than the same 64-KiB tests run without caching.

In the graph below, testing shows that an Azure NetApp Files regular volume can handle between approximately 3,500MiB/s pure sequential 256-KiB reads and approximately 2,500MiB/s pure sequential 256-KiB writes. During the tests, a 50/50 mix showed total throughput peaked higher than a pure sequential read workload.

The read-write mix for the workload was adjusted in 25% increments for each run.

Diagram of 256-KiB sequential benchmark tests.

Side-by-side comparison

To better show how caching can influence the performance benchmark tests, the following graph shows total MiB/s for 64-KiB tests with and without caching mechanisms in place. Caching provides an initial slight performance boost for total MiB/s because caching generally improves reads more so than writes. As the read/write mix changes, the total throughput without caching exceeds the results that utilize client caching.

Diagram comparing 64-KiB tests.

Parallel network connections (nconnect)

The following tests show a high I/OP benchmark using a single client with 64-KiB random workloads and a 1-TiB dataset. The workload mix generated uses a different I/O depth each time. To boost the performance for a single client workload, the nconnect mount option was leveraged for better parallelism in comparison to client mounts that didn't use the nconnect mount option. These tests were run only with caching excluded.

Results: 64 KiB, sequential, caching excluded, with and without nconnect

The following results show a scale-up test’s results when reading and writing in 4-KiB chunks on a NFSv3 mount on a single client with and without parallelization of operations (nconnect). The graphs show that as the I/O depth grows, the I/OPS also increase. But when using a standard TCP connection that provides only a single path to the storage, fewer total operations are sent per second than when a mount is able to leverage more TCP connections per mount point. In addition, the total latency for the operations is generally lower when using nconnect.

Diagram comparing 64-KiB tests without nconnect or caching.

Diagram of 64-KiB tests with nconnect but no caching.

Side-by-side comparison (with and without nconnect)

The following graphs show a side-by-side comparison of 64-KiB sequential reads and writes with and without nconnect to highlight the performance improvements seen when using nconnect: higher overall throughput, lower latency.

Diagram of comparing 64-KiB sequential reads and writes.

More information