The challenges of data proliferation and compute-intensive workloads

The rise of the Internet of Things (IoT), cloud computing, and smart phones have made it possible for businesses to harvest data from a wide range of sources and utilize it to improve their operations. Retailers can use data to track customer behavior and make their marketing more effective; manufacturers can use data to make their processes more efficient; and financial institutions can use data to detect fraud or predict market changes. As businesses gain access to new sources of data and use new technologies to analyze that data, the demand for more powerful servers will continue to grow.

Machine learning and artificial intelligence (AI) workloads have enormous potential to improve business operations, but as they gain popularity, they consume increasing amounts of processing power.² According to OpenAI, developers of ChatGPT, the computing power of their AI system doubles every 3.4 month.³ As the ML applications organizations use become more demanding, they will need more powerful servers in their data centers as well as efficient data analysis tools in the ML pipeline. Among those data analysis tools is Apache Spark^™.

Apache Spark is an open-source computing framework that converts very large data sets into smaller blocks of data for the purpose of applying machine learning algorithms and analyzing the data quickly using a distributed network of devices. For algorithms that operate on chunks of data, Spark is effective because it farms the data out to servers in the cluster, the servers process the chunks of data, and then Spark combines them for the final result. One of the main advantages of using Spark is that it can split data sets into chunks that fit in memory (when the entire data set might not) and operate with data that is entirely in memory—it doesn’t need to write to disk, which saves time. Spark is scalable: users can expand the size of their data set by adding more nodes. According to Databricks^®, Spark can process “multiple petabytes of data on clusters of over 8,000 nodes,” and Spark supports a variety of data sources, including Hadoop HDFS.⁴

We focused on two Apache Spark capabilities—k-means clustering and Bayesian classification—in our examination of the value of upgrading to the 16G Dell EMC PowerEdge R7625 server powered by 4^th Gen AMD EPYC processors along with Broadcom NICs and PERC 11 storage controllers. Using these workloads, we measured the throughput and speed of the servers. A server with better throughput and speed can process more data, handle more concurrent users, handle heavier workloads, and improve response times.

How we tested

We tested the following configurations:

• One 16G Dell PowerEdge R7625 server powered by 4^th Gen AMD EPYC 64-core processors along with Broadcom NICs and PERC 11 storage controllers
• One 15G Dell PowerEdge R7525 server powered by 3^rd Gen AMD EPYC 64-core processors along with Broadcom NICs and PERC 10 storage controllers

We configured both systems at maximum RDIMM capacity. The R7625 has a higher maximum capacity at 3TB and higher speed RAM at 4800 MT/s than the R7525 at 2TB and 3200MT/s, which is a useful upgrade for processing memory-intensive Spark workloads. We used Red Hat^® OpenShift^® virtualization. OpenShift is an open-source, Kubernetes-based container platform that offers a set of tools to manage, scale, and deploy containerized applications. For our deployment of OpenShift, we used a single-node deployment mode which is a new feature that is meant for proof of concept type environments. A typical OpenShift deployment uses three or more servers in a clustered configuration.

On each system, we created 10 OpenShift VMs with 24 cores, 96GB RAM, and one OpenShift VM with 12 cores, 32GB RAM, and one 30GB storage volume. We used this network for Spark cluster communications and Spark testing. We used Red Hat Enterprise Linux^® 8 for the OS and installed Java^™ 1.8.0, Python2^®, and Apache Maven^® 3.5.4; Apache Spark 3.0.3 with the Apache Hadoop 3.2 libraries; Apache Hadoop 3.2.4 for its HDFS capabilities; and the HiBench testing framework, version 7.1.1 with updates up to June 12, 2023 from its GitHub repository. We configured the 12-core VM as the Spark primary, and as the Hadoop manager for HDFS. We configured the remaining 10 VMs as Spark workers and Hadoop data nodes for HDFS. We used the storage volume for both the OS and for HDFS. We ran HiBench Bayes and k-means workloads from the Spark primary VM. Below is a table showing a summary of the system configurations we used in testing. For more details about our testing and configurations, read the science behind the report.

Our results

K-means clustering

For large data sets, it isn’t possible for a human to analyze the data as efficiently or effectively as a machine learning algorithm can. K-means clustering is a machine learning algorithm that aims to group similar or dissimilar data points together in clusters. By finding similarities between data points that wouldn’t be obvious with other means of analysis, k-means clustering can unlock valuable insights into individual data points, whether they are about the customers of a business, the manufacturing processes of a factory, or some other aspect of a business. These insights could help an e-commerce company offer promotions to similar types of customers or help an insurance company detect anomalies or fraud. Using the latest generation of server technology has the potential to help businesses unlock these actionable data insights faster. Tools like RapidMiner^®, ELKI, Orange, Weka^®, and MATLAB^™ rely on k-means clustering for some of types of calculations.

To better understand how upgrading server technology might benefit organizations that use k-means clustering to analyze their data, we used the HiBench benchmark suite to compare the k-means performance in terms of throughput (megabytes per second) and speed (seconds). As Figures 1 and 2 show, the new Dell PowerEdge R7625 server outperformed the previous-generation server in both measurements. The latest-generation server had 70.0 percent higher throughput and completed the k-means workload 41.2 percent faster than the previous-generation device.

These results suggest that organizations that frequently use k-means clustering to gain insights might benefit from upgrading their older servers. For an e-commerce company that provides personalized product recommendations to millions of users based on data, better throughput and faster k-means speed could allow them to tailor their recommendations more quickly. Faster throughput and speed could allow the e-commerce company to update their clustering model more frequently so that it adapts to changing customer behavior in real time. These improvements could lead to more customer engagement and higher sales.

Bayesian classification

Bayesian classification (or Bayesian inference) is a method of estimating the probability of an outcome and calculating the uncertainty around this probability using historical data. By analyzing prior outcomes, Bayesian machine learning can give organizations a statistical probability for a future outcome. A retailer may want to know the probability of a customer making a purchase after receiving a coupon code, for example. More advanced applications of Bayesian inference have helped scientists develop new drugs and assign probability to the accuracy of diagnostic tests.^11,12 Being able to quickly analyze data sets for predictions about the future can be a powerful tool for businesses and organizations.

To evaluate the Bayesian analysis performance of the servers, we used the HiBench benchmark suite to compare the total throughput, measured in megabytes per second, and the speed of analysis, in seconds. As Figure 3 shows, the 16G Dell PowerEdge R7625 achieved 19.5 percent more throughput than the previous-generation server. As Figure 4 shows, the new server was 16.3 percent faster at completing the Bayesian classification workload than the previous-generation server we compared it to.

These results indicate just how much organizations that use Bayesian machine learning to make probabilistic calculations might benefit from upgrading their aging servers. For a financial services company that uses Bayesian analysis to make investment decisions and assess risk, higher throughput and speed could allow them to handle larger data sets and run more complex models to make more accurate, real-time decisions. Alternatively, a healthcare system that uses Bayesian models for diagnosis and treatment could update patient models faster and more frequently, leading to more accurate diagnosis and better health outcomes for patients.

Performance and value – How these results can impact the bottom line

With any decision to upgrade a server environment, companies want to know that their upfront investment in new technology provides opportunities to save money further down the road. New technologies come at a price, but improvements in performance and efficiency can pay off in the long run.

Organizations can potentially save money by consolidating older servers with higher-performing, newer servers that each do more work. In our testing, a single Dell PowerEdge R7625 outperformed the Dell PowerEdge R7525 by up to 70 percent, completing 1.7 times as much k-means work as a single PowerEdge R7525. This means that two PowerEdge R7625 servers could process 3.4 times as much k-means work as one PowerEdge R7525 server. In other words, two PowerEdge R7625 servers can process the same amount of work as three PowerEdge R7525 servers with an additional 40 percent headroom. Thus, an organization that upgrades the servers in their data centers could likely reduce the total number of servers and still process the same workloads.

For each server a company can consolidate onto new gear, they can reduce their licensing cost for Red Hat OpenShift Platform Plus licensing costs for a standard 1-year subscription by $10,178.99 or by $27,820.99 for a standard 3-year subscription.^16,17 These savings don’t even take into account premium subscriptions or additional support add-ons, which would further reduce annual licensing and support costs. By reducing server counts, companies could also find savings in the reduction of cooling costs, power costs, and data center footprints. As the number of servers in a data center scales, so too do the savings associated with upgrading to the latest-generation PowerEdge R7625 servers.

Conclusion

As data proliferates and the sizes of databases grow, the potential to unlock valuable insights from them becomes increasingly dependent on fast architectures that can handle compute-intensive machine learning workloads such as k-means clustering and Bayesian inference. By upgrading to the latest servers, organizations can scale their processing power to meet the growing demands of their databases.

Larger databases and more powerful algorithms have the potential to give organizations a competitive edge. Faster servers can improve the accuracy of data-driven decisions by allowing organizations to use more complex algorithms and update ML models more frequently. To consider just two examples, improved performance could allow an e-commerce company to make better recommendations to customers and a financial services company to assess risks more accurately.

When we compared the machine learning performance of a 16G Dell PowerEdge R7625 server powered by 4^th Gen AMD EPYC 64-core processors with Broadcom NICs and PERC 11 storage controllers to a previous-generation PowerEdge server, we found performance enhancements in terms of throughput and speed, whether running k-means clustering or Bayesian workloads. These findings suggest that organizations that rely on machine learning algorithms might gain performance advantages by upgrading to the latest generation of these Dell servers.