When an entire data set is small and fits into memory (DRAM), performance is straightforward and storage system capability is less important. However, with immense data growth, a dwindling percentage of data fits into memory affordably.
Combined with the constant demand for faster and more detailed analytics, we have arrived at a data-driven crossroads: We need high performance, high capacity and affordability. Cassandra combined with enterprise SATA SSDs can help. Building with these SSDs lets us future-proof Apache Cassandra® deployments to perform soundly as active data sets grow, extending well beyond memory capacity. Cassandra’s ability to support massive scaling, combined with multi-terabyte, high-IOPS enterprise SATA SSDs, lets us build high-capacity NoSQL platforms with extreme capacity, extreme agility and extreme capability.
This technical brief highlights the performance advantages measured when we compared three four-node Cassandra clusters: one built using legacy HDDs and two built using enterprise SSDs (one per node and two per node).
Note: Due to the broad range of Cassandra deployments, we tested multiple workloads. You may find some results more relevant than others for your deployment.
- We use the terms database operations per second (OPS) and performance interchangeably in this paper.
- Each four-node enterprise SATA SSD-equipped cluster eclipsed the capability of a four-node multidrive legacy (HDD) cluster across multiple workloads and thread counts.
- The SSD configurations measured up to 31X better performance, with more consistent, lower latency.
Enterprise SSDs Meet Growing DemandsWhen we built Cassandra nodes with legacy HDD storage, we scaled out by adding more nodes to the cluster. We scaled up by upgrading to larger drives. Sometimes we did both.
Adding more legacy nodes was effective (to a point), but it quickly became unwieldy. We gained capacity and a bit more performance, but as we added to the clusters, they became larger and more complex, consuming more rack space and support resources.
Upgrading to larger HDDs was somewhat effective (also to a point) because we got more capacity per node and more capacity per cluster, but these upgrades give limited additional cluster performance.
With both techniques, performance was expensive and did not scale well with growth.
High-capacity, lightning-quick SSDs, such as the Micron® 5200 series are changing the design rules. With single SSD capacities measured in terabytes (TB), throughput in hundreds of megabytes per second (MB/s) and IOPS in tens of thousands;2 high-capacity, ultra-fast SSDs enable new design opportunities and performance thresholds.
SSD Clusters: Real Results From Immense Data SetsAs you plan your next high-capacity, high-demand Cassandra cluster, SSDs can support amazing capacity and provide compelling results. Figures 1a-1c summarize our tested storage configurations.
Figure 1a: SSD Configuration 1 - a single SSD per node
Figure 1b: SSD Configuration 2 - a pair of SSDs per node
Figure 1b: Legacy Configuration - four 15K RPM HDDs per node
- SSD Configuration 1: 1x Micron 5200 ECO (3.8TB each)
- SSD Configuration 2: 2x Micron 5200 ECO (3.8TB each)
- Legacy Configuration: 4x 15,000 RPM HDD (300GB each)
With the same number of nodes and a single SSD in each node, the 1x SSD test cluster offers a 3X capacity increase over the legacy configuration (the 2x SSD test cluster offers a 6X capacity increase). We also measured significant performance improvements across all the workloads tested with each SSD test cluster, ranging from a minimum improvement of about 1.7X to a maximum improvement of about 10.7X, along with lower and more consistent latency.
- Capacity, MB/s and IOPS vary by SSD. This paper focuses on our 3.8 TB 5200 ECO SSD. Other SSD models and/or capacities may give different results.
- We did not test YCSB workload E because it is not universally supported.
Figure 2: Relative Performance
Table 1: Workload Performance by Configuration
SSD Clusters Provide More Consistent ResponsesRead Response Consistency: Since many Cassandra deployments rely heavily on fast, consistent responses, we compared the 99th percentile read response times4 for each test cluster and workload. Figure 3 shows the 99th percentile read latency for each configuration.
An update-heavy workload, with 50% of the total I/Os writing data. At the application level, this workload is similar to recording recent session actions.
A read-mostly workload (95% read). At the application level, this workload is similar to adding metadata (such as tagging photographs or articles) to existing content.
A read-only workload. At the application level, this workload is similar to reading user profiles or other static data where profiles are constructed elsewhere.
Reading the latest entries (most recent records are the most popular). At the application level, this workload is similar to reading user status updates.
A read/modify/write workload in which records are read, changed and written back. At the application level, this workload is similar to users reading and changing data or tracking user activity.
- 99th percentile latency is the response time under which 99th of all operations complete. This is a common latency consistency measurement.
Update Response Consistency:Some Cassandra deployments are more heavily dependent on fast and consistent inserts. We also compared the 99th percentile insert response times for workload in inserts (A, B and F). Figure 4 compares the three configurations.
Figure 4: Update Response Consistency
Read/Modify/Write Response Consistency:Last, we also compared the 99th percentile read/modify/write response times for workload F (the only workload executing read/modify/write operations). Figure 5 compares the three configurations.
Figure 5: Read-Modify-Write Response Consistency
The Bottom LineHigh-capacity, high-performance SSDs can produce amazing results with Cassandra. Whether you are scaling your local or cloud-based Cassandra deployment for higher performance or faster, more consistent read responses, SSDs are a great option.
We tested two clusters for database performance and read responsiveness across multiple workloads. We built a legacy cluster using four 300GB 15K RPM HDDs (RAID 0) in each node and two SSD clusters, one with a single 3.8TB Micron 5200 ECO and a second with a pair of them in each node. The results were amazing.
The single SSD per node test cluster showed a significant increase in performance over all the workloads, ranging from a low of about 1.7X up to a high of 7.5X compared to the legacy configuration. The two SSD per node test cluster was still more impressive with workload increases ranging from 2.2X to 10.7X compared to the legacy configuration. We also found that the SSD-based clusters’ read responses were much faster with far greater consistency.
We expect great performance when our data set fits into memory, but immense data growth means that smaller and smaller portions of that data fit into memory affordably.
We are at a crossroads. Business demands drive us toward higher performance, and data growth drives us toward affordable capacity. When we combine these, the answer is clear: Enterprise SSDs deliver Cassandra strong results, helping tame performance demands and data growth.
Given data like the above, our customers are increasingly finding that deploying SSDs in the data center is a high-value option for better overall total cost of ownership (TCO). If you are curious how this compares to other configurations, try Micron’s Move2SSD TCO Tool to estimate the savings you can see from deploying SSDs versus existing architecture.
How We TestedTable 1 shows the tested configurations, the types of storage devices used, and the number and capacity of each as well as the number of nodes in each Cassandra test cluster. Table 2 shows the hardware and software configuration parameters used.
Table 1: Tested Configurations
Table 2: Configuration Parameters
Our test methodology approximates real-world deployments and uses for a Cassandra database. Although the test configuration is relatively small (four nodes in each cluster), Cassandra’s scaling technology means these results are also relevant to larger deployments.
- Four nodes host the database.
- The replication factor for the database was set to 3 (there are three copies of the data and the cluster can sustain the loss of two data nodes and continue to function) creating a 1.5TB database.
Table 3 shows the percentage of data owned by each of the four nodes.
Table 3: Data Distribution Across Nodes
Table 4: Test Parameters
Table 5 shows the IO profiles for tested YCSB workloads (additional details are available at YCSB Core Workloads).
Table 5: Workloads
To download a pdf version of this Technical Brief, click here