This page describes a series of tests for measuring performance after the installation of the Weka system. The same tests can be used to test the performance of any other storage solution.
About Weka performance tests
There are three main performance metrics when measuring a storage system performance:
Latency, which is the time from operation initiation to completion
The number of different IO operations (read/write/metadata) that the system can process concurrently
The bandwidth of data that the system can process concurrently
Each of these performance metrics applies to read operations, write operations, or a mixture of read and write operations.
When measuring the Weka system performance, different mount modes produce different performance characteristics. Additionally, client network configuration (using either user-space DPDK networking or kernel UDP) also significantly affects performance.
Note: All performance tests listed here are generic and not specific to the Weka system. They can be used to compare the Weka storage system to other storage systems or a local storage device.
Note: There is a difference between single client performance to aggregated performance. When running the tests listed below from one client, the client will limit the test's performance. In general, several clients will be required to maximize the performance of a Weka cluster.
The FIO tool
The FIO tool is a generic open-source storage performance testing tool that can be defined as described here. In this documentation, the usage of FIO version 3.20 is assumed.
All FIO testing is done using the client/server capabilities of FIO. This makes multiple client testing easier since FIO reports aggregated results for all clients under the test. Single client tests are run the same way to keep the results consistent.
Start the FIO server on every one of the clients:
fio --server --daemonize=/tmp/fio.pid
Run the test command from one of the clients, note, the clients need to be mounted to a Weka filesystem.
An example of launching a test (sometest) on all clients in a file (clients.txt) using the server/client model:
fio --client=clients.txt sometest.txt
An example for the clients' file, when running multiple clients:
The single-client or aggregated tests deffer in the clients participating in the test, as defined in the clients.txt.
MDTest
MDTest is a generic open-source metadata performance testing tool. In this documentation, the usage of version 1.9.3 is assumed.
MDTest uses an MPI framework to coordinate the job across multiple nodes. The results presented here were generated using the MPICH version 3.3.2 and can be defined as described here. While it's possible to have variations with different MPI versions, most are based on the same ROMIO and will perform similarly.
Weka client performance tests
Overall, the tests contained on this page are designed to show off the sustainable peak performance of the filesystem. Care has been taken to make sure they are realistic and reproducible.
Where possible, the benchmarks try to negate the effects of caching. For file testing, o_direct calls are used to bypass the client's cache. In the case of metadata testing, each phase of testing uses different clients. Also, between each test, the Linux caches are flushed to ensure all data being accessed is not present in the cache. While applications will often take advantage of cached data and metadata, this testing focuses on the filesystem's ability to deliver data independent of caching on the client.
While we provide below the output of one iteration, we ran each test several times and provided the average results in the Result Summary.
Note: If the client uses a 100 Gbps NIC or above, mounting the Weka filesystem with more than one core is required to maximize client throughput.
Test read throughput
This test measures the client throughput for large (1MB) reads. The job below tries to maximize the read throughput from a single client. The test utilizes multiple threads, each one performing 1 MB reads.
In this test output example, results show a bandwidth of 8.95 GiB/s from a single client.
Test write throughput
This test measures the client throughput for large (1MB) writes. The job below tries to maximize the write throughput from a single client. The test utilizes multiple threads, each one performing 1MB writes.
In this test output example, results show a bandwidth of 6.87 GiB/s.
Test read IOPS
This test measures the ability of the client to deliver concurrent 4KB reads. The job below tries to maximize the system read IOPS from a single client. The test utilizes multiple threads, each one performing 4KB reads.
In this test output example, results show 390,494 IOPS from a single client.
Test write IOPS
This test measures the ability of the client to deliver concurrent 4KB writes. The job below tries to maximize the system write IOPS from a single client. The test utilizes multiple threads, each one performing 4KB writes.
In this test output example, results show 288,215 IOPS from a single client.
Test read latency
This test measures the minimal achievable read latency under a light load. The test measures the latency over a single-threaded sequence of 4KB reads across multiple files. Each read is executed only after the previous read has been served.
In this test output example, results show an average latency of 229 microseconds, where 99.5% of the writes terminated in 334 microseconds or less.
Test write latency
This test measures the minimal achievable write latency under a light load. The test measures the latency over a single-threaded sequence of 4KB writes across multiple files. Each write is executed only after the previous write has been served.
In this test output example, results show an average latency of 226 microseconds, where 99.5% of the writes terminated in 293 microseconds or less.
Test metadata performance
The test measures the rate of metadata operations (such as create, stat, delete) across the cluster. The test uses 20 million files: it uses 8 client hosts, and multiple threads per client are utilized (136), where each thread handles 18382 files. It is invoked 3 times and provides a summary of the iterations.
