Category: Performance Tuning

Reitse Eskens continues a series on performance testing Azure SQL DB tiers:

So far, my blogs have been on the different Azure SQL DB offerings where there are differences between DTU and CPU based. But in general, the design is recognizable. With the hyperscale tier, many things change. There are still cores and memory of course, but the rest of the design is totally different. I won’t go into all the details, you’re better off reading them here [https://learn.microsoft.com/en-us/azure/azure-sql/database/service-tier-hyperscale?view=azuresql] and here [https://learn.microsoft.com/en-us/azure/azure-sql/database/hyperscale-architecture?view=azuresql] , but the main differences are the support of up to 100 TB of data in one database (all the other tiers max out at 40 TB), fast database restores based on file snapshots, rapid scale out and rapid scale up.

There are differences in testing this one versus the others, so buyer beware.

Reitse Eskens continues a series on comparing the performance of different Azure SQL DB tiers:

The compute optimized tier starts at 8 cores as a minimum and goes up to 72.
The maximum storage starts at 1 TB and  goes up to 4 TB. The number of TempDB files is undisclosed but the size starts at 37 GB and goes up to 333 GB. A huge difference with the regular provisioned one! The disk limitations are disclosed as well. Log rate starts at 36 Mbps and maxing out at 50. The same goes for the data; starting at 2560 iops and maxing out at 12.800. One iop (or Input Output oPeration) is connected to the disk cluster size. As these are 4 kb, reading or writing one data page (8Kb) equals two iops. 2560 Iops equals something of 10 MB per second. The top end goes to about 50 MB per second. Keep this in mind when you’re working out which tier you need, because usually disk performance can be essential.

Read on to see how it performs on Reitse’s standardized test workload.

Reitse Eskens doesn’t need a server:

In my previous blog I wrote about the premium tier, the one that can be compared with the business critical tier. Now we’re moving away from the DTU models and back to what we DBA’ers really understand, cores, memory and disks. Before I’m going to dive into the limitations, there’s one thing you need to understand. The serverless tier is made for intermittent use. If you’re using the tier for more than 25% of the time (or about 183 hours per month), you’re better of going provisioned. This has nothing to do with performance but everything with cost. The tipping point of provisioned being cheaper is around 25% of the time.

There’s some solid advice on how to get the service to go to sleep but the bulk of the article revolves around performance.

Reitse Eskens is moving on up:

The standard tier starts at 125 DTU’s and goes up to 4000. DTU’s are made up from a magic mix of CPU, memory, read iops and write iops. An iop (Input Output oPeration) should be a 4kb (disk cluster size) read or write. 125 DTU translates to 500 Kb/sec up to 32.000 Kb/sec. As we’re used to datapages which are 8Kb in size, you could say these databases are able to pull 62 to 4.000 pages per second from disk. When there are simultaneous writes, you’ll share the performance. At least that’s my interpretation of the IOP. For the DTU part, I’m still struggling to get a good grip on what it exactly is, beyond the magic mix.

It’s also a good idea to compare this to what the Standard tier has to offer. The general data patterns look similar with respect to elbows but the magnitudes are quite different, with Premium P1 starting out around Standard S4 in the test for insertion but more like S3 for selects.

Reitse Eskens continues a series on performance comparisons for Azure SQL DB:

This tier is more expensive than the basic, but starting at 12 euro’s per month up to 3723 euro’s you have a wider range of spending your money and with that a wider performance range. The standard tier is suited for general purpose workloads and can be compared with the general purpose tier whereas the latter works with cores. Standard tier works with DTU’s. The concept of a DTU is a difficult one as the documentation states it’s a blend of CPU, Memory, reads and writes. If you hit a limitation, you’ll be throttled. You can read more about the DTU model here.

One thing I wish Reitse had done in the images was to show them in log scale—there’s a consistent L curve for each (which is good) but makes it hard to see anything after about Standard S4.