# Benchmark Cassandra performance ## In this learning path - [Introduction](https://learn.arm.com/learning-paths/servers-and-cloud-computing/cassandra-on-gcp/) - [Get started with Cassandra on Google Axion C4A](https://learn.arm.com/learning-paths/servers-and-cloud-computing/cassandra-on-gcp/background/) - [Create a Google Axion C4A Arm virtual machine](https://learn.arm.com/learning-paths/servers-and-cloud-computing/cassandra-on-gcp/instance/) - [Install Apache Cassandra](https://learn.arm.com/learning-paths/servers-and-cloud-computing/cassandra-on-gcp/installation/) - [Test Cassandra baseline functionality](https://learn.arm.com/learning-paths/servers-and-cloud-computing/cassandra-on-gcp/baseline/) - [Benchmark Cassandra performance](https://learn.arm.com/learning-paths/servers-and-cloud-computing/cassandra-on-gcp/benchmarking/) - [Next Steps](https://learn.arm.com/learning-paths/servers-and-cloud-computing/cassandra-on-gcp/_next-steps/) ## Benchmark Cassandra with cassandra-stress You can perform Cassandra benchmarking using the built-in `cassandra-stress` tool, which measures database performance under different workloads such as write, read, and mixed operations. ## Verify cassandra-stress installation Cassandra comes with a built-in tool called `cassandra-stress` that is used for testing performance. It’s located in the `tools/bin/` folder of your Cassandra installation. ``` ls ~/cassandra/tools/bin | grep cassandra-stress ``` If you see `cassandra-stress` in the list, the tool is installed and ready to use. Check the tool’s help options to verify it works correctly: ``` ~/cassandra/tools/bin/cassandra-stress help ``` The output is similar to: ``` __output__ Usage: cassandra-stress [options] __output__ Help usage: cassandra-stress help __output__ ---Commands--- read : Multiple concurrent reads - the cluster must first be populated by a write test write : Multiple concurrent writes against the cluster mixed : Interleaving of any basic commands, with configurable ratio and distribution - the cluster must first be populated by a write test counter_write : Multiple concurrent updates of counters. counter_read : Multiple concurrent reads of counters. The cluster must first be populated by a counterwrite test. user : Interleaving of user provided queries, with configurable ratio and distribution help : Print help for a command or option print : Inspect the output of a distribution definition version : Print the version of cassandra stress ``` The list of commands and options confirms that your setup is correct and you’re ready to start testing Cassandra’s performance. ## Run a basic write test Insert 10,000 rows with 50 concurrent threads using `cassandra-stress`: ``` ~/cassandra/tools/bin/cassandra-stress write n=10000 -rate threads=50 ``` - **write** → Performs only write operations on the Cassandra cluster. - **n=10000** → Specifies the number of rows to insert during the benchmark test. - **-rate threads=50** → Sets the number of concurrent worker threads simulating multiple clients writing to the cluster. The output is similar to: ``` __output__ ******************** Stress Settings ******************** __output__ Command: __output__ Type: write __output__ Count: 10,000 __output__ No Warmup: false __output__ Consistency Level: LOCAL_ONE __output__ Target Uncertainty: not applicable __output__ Key Size (bytes): 10 __output__ Counter Increment Distibution: add=fixed(1) __output__ Rate: __output__ Auto: false __output__ Thread Count: 50 __output__ OpsPer Sec: 0 __output__ Population: __output__ Sequence: 1..10000 __output__ Order: ARBITRARY __output__ Wrap: true __output__ Insert: __output__ Revisits: Uniform: min=1,max=1000000 __output__ Visits: Fixed: key=1 __output__ Row Population Ratio: Ratio: divisor=1.000000;delegate=Fixed: key=1 __output__ Batch Type: not batching __output__ Columns: __output__ Max Columns Per Key: 5 __output__ Column Names: [C0, C1, C2, C3, C4] __output__ Comparator: AsciiType __output__ Timestamp: null __output__ Variable Column Count: false __output__ Slice: false __output__ Size Distribution: Fixed: key=34 __output__ Count Distribution: Fixed: key=5 __output__ Errors: __output__ Ignore: false __output__ Tries: 10 __output__ Log: __output__ No Summary: false __output__ No Settings: false __output__ File: null __output__ Interval Millis: 1000 __output__ Level: NORMAL __output__ Mode: __output__ API: JAVA_DRIVER_NATIVE __output__ Connection Style: CQL_PREPARED __output__ Protocol Version: V5 __output__ Username: null __output__ Password: null __output__ Max Pending Per Connection: 128 __output__ Connections Per Host: 8 __output__ Compression: NONE __output__ Node: __output__ Nodes: [localhost] __output__ Is White List: false __output__ Datacenter: null __output__ Schema: __output__ Keyspace: keyspace1 __output__ Replication Strategy: org.apache.cassandra.locator.SimpleStrategy __output__ Replication Strategy Options: {replication_factor=1} __output__ Table Compression: null __output__ Table Compaction Strategy: null __output__ Table Compaction Strategy Options: {} __output__ Transport: __output__ Truststore: null __output__ Truststore Password: null __output__ Keystore: null __output__ Keystore Password: null __output__ SSL Protocol: TLS __output__ SSL Algorithm: null __output__ SSL Ciphers: TLS_RSA_WITH_AES_128_CBC_SHA,TLS_RSA_WITH_AES_256_CBC_SHA __output__ Port: __output__ Native Port: 9042 __output__ JMX Port: 7199 __output__ JMX: __output__ Username: null __output__ Password: *not set* __output__ Graph: __output__ File: null __output__ Revision: unknown __output__ Title: null __output__ Operation: WRITE __output__ TokenRange: __output__ Wrap: false __output__ Split Factor: 1 __output__ Credentials file: __output__ File: *not set* __output__ CQL username: *not set* __output__ CQL password: *not set* __output__ JMX username: *not set* __output__ JMX password: *not set* __output__ Transport truststore password: *not set* __output__ Transport keystore password: *not set* __output__ Reporting: __output__ Output frequency: 1s __output__ Header frequency: *not set* __output__ Connected to cluster: Test Cluster, max pending requests per connection 128, max connections per host 8 __output__ Datacenter: datacenter1; Host: localhost/127.0.0.1:9042; Rack: rack1 __output__ Created keyspaces. Sleeping 1s for propagation. __output__ Sleeping 2s... __output__ Warming up WRITE with 2500 iterations... __output__ Running WRITE with 50 threads for 10000 iteration __output__ type total ops, op/s, pk/s, row/s, mean, med, .95, .99, .999, max, time, stderr, errors, gc: #, max ms, sum ms, sdv ms, mb __output__ total, 10000, 10690, 10690, 10690, 3.7, 2.8, 9.5, 16.7, 28.9, 38.4, 0.9, 0.00000, 0, 0, 0, 0, 0, 0 __output__ Results: __output__ Op rate : 10,690 op/s [WRITE: 10,690 op/s] __output__ Partition rate : 10,690 pk/s [WRITE: 10,690 pk/s] __output__ Row rate : 10,690 row/s [WRITE: 10,690 row/s] __output__ Latency mean : 3.7 ms [WRITE: 3.7 ms] __output__ Latency median : 2.8 ms [WRITE: 2.8 ms] __output__ Latency 95th percentile : 9.5 ms [WRITE: 9.5 ms] __output__ Latency 99th percentile : 16.7 ms [WRITE: 16.7 ms] __output__ Latency 99.9th percentile : 28.9 ms [WRITE: 28.9 ms] __output__ Latency max : 38.4 ms [WRITE: 38.4 ms] __output__ Total partitions : 10,000 [WRITE: 10,000] __output__ Total errors : 0 [WRITE: 0] __output__ Total GC count : 0 __output__ Total GC memory : 0 B __output__ Total GC time : 0.0 seconds __output__ Avg GC time : NaN ms __output__ StdDev GC time : 0.0 ms __output__ Total operation time : 00:00:00 __output__ END ``` ## Run a read test Run a read benchmark on your Cassandra database using `cassandra-stress`. This simulates multiple clients reading from the cluster at the same time and records performance metrics such as throughput and latency. ``` ~/cassandra/tools/bin/cassandra-stress read n=10000 -rate threads=50 ``` The output is similar to: ``` __output__ ******************** Stress Settings ******************** __output__ Command: __output__ Type: read __output__ Count: 10,000 __output__ No Warmup: false __output__ Consistency Level: LOCAL_ONE __output__ Target Uncertainty: not applicable __output__ Key Size (bytes): 10 __output__ Counter Increment Distibution: add=fixed(1) __output__ Rate: __output__ Auto: false __output__ Thread Count: 50 __output__ OpsPer Sec: 0 __output__ Population: __output__ Distribution: Gaussian: min=1,max=10000,mean=5000.500000,stdev=1666.500000 __output__ Order: ARBITRARY __output__ Wrap: false __output__ Insert: __output__ Revisits: Uniform: min=1,max=1000000 __output__ Visits: Fixed: key=1 __output__ Row Population Ratio: Ratio: divisor=1.000000;delegate=Fixed: key=1 __output__ Batch Type: not batching __output__ Columns: __output__ Max Columns Per Key: 5 __output__ Column Names: [C0, C1, C2, C3, C4] __output__ Comparator: AsciiType __output__ Timestamp: null __output__ Variable Column Count: false __output__ Slice: false __output__ Size Distribution: Fixed: key=34 __output__ Count Distribution: Fixed: key=5 __output__ Errors: __output__ Ignore: false __output__ Tries: 10 __output__ Log: __output__ No Summary: false __output__ No Settings: false __output__ File: null __output__ Interval Millis: 1000 __output__ Level: NORMAL __output__ Mode: __output__ API: JAVA_DRIVER_NATIVE __output__ Connection Style: CQL_PREPARED __output__ Protocol Version: V5 __output__ Username: null __output__ Password: null __output__ Max Pending Per Connection: 128 __output__ Connections Per Host: 8 __output__ Compression: NONE __output__ Node: __output__ Nodes: [localhost] __output__ Is White List: false __output__ Datacenter: null __output__ Schema: __output__ Keyspace: keyspace1 __output__ Replication Strategy: org.apache.cassandra.locator.SimpleStrategy __output__ Replication Strategy Options: {replication_factor=1} __output__ Table Compression: null __output__ Table Compaction Strategy: null __output__ Table Compaction Strategy Options: {} __output__ Transport: __output__ Truststore: null __output__ Truststore Password: null __output__ Keystore: null __output__ Keystore Password: null __output__ SSL Protocol: TLS __output__ SSL Algorithm: null __output__ SSL Ciphers: TLS_RSA_WITH_AES_128_CBC_SHA,TLS_RSA_WITH_AES_256_CBC_SHA __output__ Port: __output__ Native Port: 9042 __output__ JMX Port: 7199 __output__ JMX: __output__ Username: null __output__ Password: *not set* __output__ Graph: __output__ File: null __output__ Revision: unknown __output__ Title: null __output__ Operation: READ __output__ TokenRange: __output__ Wrap: false __output__ Split Factor: 1 __output__ Credentials file: __output__ File: *not set* __output__ CQL username: *not set* __output__ CQL password: *not set* __output__ JMX username: *not set* __output__ JMX password: *not set* __output__ Transport truststore password: *not set* __output__ Transport keystore password: *not set* __output__ Reporting: __output__ Output frequency: 1s __output__ Header frequency: *not set* __output__ Sleeping 2s... __output__ Warming up READ with 2500 iterations... __output__ Connected to cluster: Test Cluster, max pending requests per connection 128, max connections per host 8 __output__ Datacenter: datacenter1; Host: localhost/127.0.0.1:9042; Rack: rack1 __output__ Running READ with 50 threads for 10000 iteration __output__ type total ops, op/s, pk/s, row/s, mean, med, .95, .99, .999, max, time, stderr, errors, gc: #, max ms, sum ms, sdv ms, mb __output__ total, 1540, 1540, 1540, 1540, 8.1, 6.2, 19.2, 38.4, 73.3, 80.9, 1.0, 0.00000, 0, 0, 0, 0, 0, 0 __output__ total, 9935, 8395, 8395, 8395, 5.9, 4.2, 16.7, 33.1, 57.3, 86.0, 2.0, 0.48892, 0, 0, 0, 0, 0, 0 __output__ total, 10000, 4217, 4217, 4217, 8.5, 4.2, 27.1, 27.4, 27.4, 27.4, 2.0, 1.89747, 0, 0, 0, 0, 0, 0 __output__ Results: __output__ Op rate : 4,962 op/s [READ: 4,962 op/s] __output__ Partition rate : 4,962 pk/s [READ: 4,962 pk/s] __output__ Row rate : 4,962 row/s [READ: 4,962 row/s] __output__ Latency mean : 6.3 ms [READ: 6.3 ms] __output__ Latency median : 4.5 ms [READ: 4.5 ms] __output__ Latency 95th percentile : 17.4 ms [READ: 17.4 ms] __output__ Latency 99th percentile : 33.4 ms [READ: 33.4 ms] __output__ Latency 99.9th percentile : 59.6 ms [READ: 59.6 ms] __output__ Latency max : 86.0 ms [READ: 86.0 ms] __output__ Total partitions : 10,000 [READ: 10,000] __output__ Total errors : 0 [READ: 0] __output__ Total GC count : 0 __output__ Total GC memory : 0 B __output__ Total GC time : 0.0 seconds __output__ Avg GC time : NaN ms __output__ StdDev GC time : 0.0 ms __output__ Total operation time : 00:00:02 __output__ END ``` ## Understand benchmark results The metrics below explain what each value in the `cassandra-stress` output represents: - Op rate (operations per second): the number of read operations Cassandra successfully executed per second. - Partition rate: the number of partitions read per second. Since this is a read test, the partition rate equals the op rate. - Row rate: the number of rows read per second. Again, for this test it equals the op rate. - Latency mean: the average time taken for each read request to complete. - Latency median: the 50th percentile latency - half of the operations completed faster than this time. - Latency max: the slowest single read request during the test. - Total partitions: the total number of partitions read during the test. - Total errors: the number of failed read operations. - GC metrics (Garbage Collection): shows whether JVM garbage collection paused Cassandra during the test. - Total operation time: the total wall-clock time taken to run the benchmark. ## Benchmark summary for Arm64 Results from the run on the `c4a-standard-4` (4 vCPU, 16 GB memory) Arm64 VM in GCP (SUSE shown; Ubuntu results were very similar): | Metric | Write Test | Read Test | |-----------------------------|------------|-----------| | Operation Rate (op/s) | 10,690 | 4,962 | | Partition Rate (pk/s) | 10,690 | 4,962 | | Row Rate (row/s) | 10,690 | 4,962 | | Latency Mean | 3.7 ms | 6.3 ms | | Latency Median | 2.8 ms | 4.5 ms | | Latency 95th Percentile | 9.5 ms | 17.4 ms | | Latency 99th Percentile | 16.7 ms | 33.4 ms | | Latency 99.9th Percentile | 28.9 ms | 59.6 ms | | Latency Max | 38.4 ms | 86.0 ms | | Total Partitions | 10,000 | 10,000 | | Total Errors | 0 | 0 | | Total GC Count | 0 | 0 | | Total GC Memory | 0 B | 0 B | | Total GC Time | 0.0 s | 0.0 s | | Total Operation Time | 0:00:00 | 0:00:02 | ## What you’ve accomplished and what’s next You’ve successfully deployed Apache Cassandra 5.0.5 on a Google Axion C4A Arm-based virtual machine, validated its functionality, and measured its performance using cassandra-stress. The benchmark results on Google Axion C4A Arm-based instances demonstrate strong performance characteristics. Write operations achieved high throughput of 10,690 op/s, while read operations reached 4,962 op/s on the `c4a-standard-4` Arm64 VM. Write latency was notably low with a mean of 3.7 ms compared to reads at 6.3 ms, indicating fast write processing on this Arm64 VM. The 95th and 99th percentile latencies show consistent performance, with writes significantly faster than reads. Zero errors or GC overhead confirm stable and reliable benchmarking results. The Arm64 VM provides efficient and predictable performance, making it suitable for high-throughput Cassandra workloads. The low write latencies and high operation rates demonstrate that Arm-based infrastructure can effectively handle database operations that require both speed and consistency. These results provide a solid baseline for evaluating Cassandra performance on Arm64 architecture and can guide decisions about instance sizing and configuration for production deployments. To continue building on this foundation, you can explore advanced Cassandra configurations such as multi-node cluster deployments, replication strategies for high availability, or performance tuning for specific workload patterns. You might also investigate integrating Cassandra with application frameworks or comparing performance across different Arm-based instance types to optimize for your use case.