Understanding Azure Virtual Machine IOPS, throughput and disk latency – Part 2

Continuing on from Part 1 : https://blogs.technet.microsoft.com/andrewc/2016/09/09/understanding-azure-virtual-machine-iops-throughput-and-disk-latency

Part 1 showed the IOPS, throughout and latency characteristics of an Azure disk when there was no caching involved.

This post covers the impact of caching and premium storage.

Test 4 – Read Cache – Standard Disk.

We will re-run the tests using the F: drive, which is a standard disk with Read Caching turned on

Results :

Read IOPS is now showing ~32000 – on a standard 500 IOPS disk.

Using the same F: drive, change the Access Specification to 100% write and re-run.

In this case, the IOPS are still ~500 as write operations are not cached.

Test 5 – Read/Write Cache – Standard Disk

Now switch to drive G: which has both read and write caching enabled.

With the access specification still set to 100% write, we find ~32000 IOPS

Summary:

• Caching *may* make significant difference to the apparent performance of the disk subsystem
  
  • I say apparent, because actual throughput to/from the disk subsystem is unchanged – but cache can reduce the number of IOs that occur.
    
• Read cache will not necessarily improve performance for any given application – it depends on the application and data profile.
  
• Write cache should be used with care as data loss is possible in the event of a hardware issue with the physical server holding the cache before it is written to disk.
  
• Consult any guidance or best practices for the specific application in question
  

Test 6 – SSD Premium Storage

Switching to drive I: (Premium P30 disk) and rerunning the same tests, we see:

Read operations when using a small block size generates 5000 IOPS as per spec.

Changing the block size to 256k and re-running :

The IOPS is lower because we hit the 200MB/s throughput limit:

Testing latency by changing outstanding IO to 1

Latency is ~1.6ms (compared to 5ms + for standard storage)

Summary:

• Premium storage can achieve 10 times the IOPS, 3 times the throughput and 1/3 the latency when compared to standard disks.
  
• While the raw numbers are better, the same constraints around single threaded IO still apply.
  

Test 7 – Using the Local D: temp disk

All Azure virtual machines have a temporary disk, presented as D: drive on Windows and /dev/sdb on Linux.

The temp disk is designed for temporary working storage for applications and processes, but not long term storage. The contents will be lost if the VM is moved to another host.

Re-running the previous tests :

Read Performance small blocks:

Read Performance large blocks:

IO Latency when using outstanding IO set 1:

Summary

• The local temporary disk can provide high IOPS, high throughput and low latency compared to permanent blob/vhd storage.
• Use with care as the data is transient.
• Using D: for SQL Server tempdb is supported and documented :
  • https://blogs.technet.microsoft.com/dataplatforminsider/2014/09/25/using-ssds-in-azure-vms-to-store-sql-server-tempdb-and-buffer-pool-extensions/
  • https://azure.microsoft.com/en-gb/documentation/articles/virtual-machines-windows-classic-sql-perf/