Google Cloud's C3D Instances Provide Strong Performance Value For PingCAP's TiDB

Some time ago I ran through a number of benchmarks of Google Cloud's C3D VMs powered by AMD EPYC Genoa processors. The AMD EPYC 9004 series showed terrific performance with strong generational improvements over the Intel Xeon Scalable processors. Following that a request came in to examine the PingCAP TiDB database performance given its growing popularity. In this article we'll review those benchmarks showing how GCE C3D delivers strong performance advantages for TiDB.

For those unaware, TiDB is an open-source, distributed SQL database with built-in horizontal scalability across nodes and native support for OLTP and OLAP workloads. Developed by PingCAP, TiDB also guarantees auto-failover and self-healing while providing automatic sharding. You can learn more about this database at PingCAP.com.

Since TiDB offers MySQL compatibility, sysbench was used for facilitating benchmarks. Beyond the TiDB cloud benchmarks shared today, I've been running TiDB locally more and more across various scales of systems. You can find more of my local CPU results for TiDB via this test profile page. Additionally, PingCAP has also developed Web3Bench as another database benchmark more catered to TiDB "Web3" workloads. This benchmark will be looked at in the future on Phoronix.

The Google Cloud instances tested for this comparison included:

c2-standard-60: The top-end C2 machine type that offers 60 vCPUs with 240GB of memory for $3.1321 USD per hour. Google Cloud C2 VMs are powered by 2nd Generation Intel Xeon Scalable (Cascade Lake) processors.

c2d-standard-56: The 56 vCPU instance with 224GB of memory for these GCE VMs focused on HPC workloads. C2D instances are powered by prior generation AMD EPYC Milan processors. Pricing is at $2.5422 per hour.

c3d-standard-30: The 30 vCPU instance of the new C3D series powered by AMD EPYC Genoa. The c3d-standard-30 has 120GB of RAM and is priced at $1.362 USD per hour.

c3d-standard-60: The 60 vCPU instance of the EPYC Genoa C3D series with 240GB of RAM. The c3d-standard-60 during testing was priced at $2.724 per hour.

n2d-standard-64: The general purpose N2D instance powered by AMD EPYC Rome with 64 vCPUs and 256GB of RAM that is priced at $2.703744 per hour.

t2a-standard-48: The highest-tier, 48 vCPU offering for Tau T2A Ampere powered instances with 48 phyysical cores. Current pricing at $1.848 per hour.

t2d-standard-60: The Tau T2A instance powered by AMD EPYC Milan with 60 vCPUs and 240GB of RAM. This 60 CPU Tau T2A instance is currently priced at $2.53476 per hour.

For each of the instances tested they used the instance size as close as possible to 60 vCPUs -- with some instance families instead being sized for 56 vCPUs or Tau T2A tapping out at 48 vCPUs. Plus with the C3D is also a 30 vCPU run for reference to help with analyzing scaling of the given benchmark. The C3 instance type unfortunately wasn't tested since the C3 instance sizes don't align with the other instance sizes tested for this article.

For these TiDB benchmarks, each Google Compute Engine VM benchmark was tested with a 215GB SSD persistent disk using Ubuntu 22.04 LTS with the Linux 6.2 kernel atop EXT4. The on-demand hourly rates were used when calculating the performance-per-dollar of these benchmarks. AMD kindly provided the cloud computing credits to cover the cost of this independent public cloud benchmarking.