In our previous post, we shared our measurements of the performance of the hypervisor after installing patches against the vulnerabilities of Meltdown and Specter. Today it's time to talk about the performance of the data warehouse.
Thanks to optimizations of VzKernel and its recompilation with the “Retpoline” option, we replaced vulnerable machine code sequences and were able to almost completely eliminate performance problems caused by the need to protect the hypervisor from vulnerabilities in Intel processors. As a result, the decrease in productivity was reduced to 1-2%. However, against this background, many have questions about the work of the data warehouse. And this is not surprising, because in hyperconvergent environments, distributed data storage plays a fundamental role, and with slow storage, all the performance benefits of virtual machines and containers can disappear.
Today we want to share with you two demonstrative tests that were carried out to assess the performance of virtual machines and the density of data placement in a distributed storage VZ Storage, which is built into the Virtuozzo 7 product family. A test cluster was a cluster of 6 nodes, with data storage directly only 5 of them were occupied, and virtual machines were located on the remaining node.
Each node had the following configuration:
- CPU: 2 x Intel Xeon E5-2620 v4 @ 2.1 GHz
- RAM: 64 GB DDR4 2133 MHz
- NET: 2 x 10 Gbps - two separate subnets for dividing test traffic and distributed data storage traffic
- Capacities:
- HDD: 8 x 1 TB 7200 RPM - including 7 HDD for chunk servers (CS)
- SSD: 400 GB Intel DC P3600 PCIe - for metadata (MDS), logging and client cache
One of the disks on each node was allocated for the system, and the remaining 7 - for the chank-server (CS) for data storage. As a result, 42 cluster servers turned out in the cluster. To manage this farm, we launched 3 metadata servers (MDS). Data replication was implemented using a 3: 2 scheme, which can be considered a standard solution for most typical tasks.
According to the results of the WebBench test, where we evaluated the performance and density of virtual machines running Windows Server 2012 R2, the number of requests for virtual storage in VZ7 is much higher, and the overall performance by an average of 30% exceeds the results of the previous generation of storage that came with VZ6 . At the same time, VZ Storage together with the Virtuozzo 7 hypervisor can support simultaneous operation of over 100 virtual machines on a cluster of this size, ensuring them acceptable performance.
WebBench: VMtorage VM Windows 2012 R2 VM Density
The second test was conducted using the SysBench utility and emulated no longer web requests, but OLTP transactions. We loaded the same virtual machines with Microsoft Windows Server 2012 R2 on the same cluster and got even more interesting results. In addition to performance advantages with the number of VMs from 30 pieces, VZ7 showed a higher density of accommodation, coping with the simultaneous operation of more than 100 virtual machines. At the same time, the outdated storage on the VZ6 demonstrates acceptable performance for up to 60 virtual machines on a cluster of reduced size.
SysBench: VMtorage VM 2012 R2 VM Density
And a little more about Erasure Coding
In addition to the above, Virtuozzo remains committed to using compression technologies based on Reed-Solomon codes or Erasure Coding. Despite a broad discussion of this technology, many still prefer to use direct copies and store up to 3 copies of their data in their network. However, as practice has shown, this approach reduces network performance and slows down backup processes.
To verify this fact, we collected two clusters, 6 nodes each. On both clusters, 3 metadata servers (MDS) and 66 data chank-servers (CS) were running on top of SAS 15K disks. One cluster was used to host the virtual machines, and the other was used for backup. We tried two ways of placing backups: EC in 3 + 2 mode (for three pieces of data - two hash sums) and full backup 3: 2 (two full copies of data are stored in the network). From the point of view of data protection, the configurations are identical - they make it possible to recover all the information even if two points of failure occur. However, in terms of performance, the EC shows much better results.
Erasure Coding and data replication in a parallel VM backup scenario
The abscissa shows the number of virtual machines that are simultaneously involved in the backup process. And the ordinate is the average backup creation speed in MB / s. The speed is calculated per node, so that the overall throughput and performance of the cluster is much higher, a multiple of the number of nodes. From the graph it can be seen that while simultaneously backing up 15 virtual machines from each node, the performance gain due to the use of EC is at the level of 10%.
Conclusion
These tests show the advantages of an updated architecture and improved VZ Storage performance when working with MS Windows virtual machines, which are traditionally heavier optimized and compacted than VMs with Linux guest files that can be converted to system containers in general. In this case, we used SAS 15K hard drives in the test, rather than solid-state drives, for which the results would be even higher due to an increase in the overall response time and the speed of the storage subsystem.