Laboratory IT equipment budgets are actually extremely limited. Although the cost of using a single all-flash array is relatively high, it really can reduce the enormous cost of building compute nodes. It can also significantly reduce future operating costs, including electricity bills. Combined with its reduced computing times, it can assist laboratories in performing more analysis more accurately.– AccelStor Senior Product Manager Alex Ho
• The overall computing time of a traditional hard disk array was compared to that of the P310 all-flash array under the same structural model.
The P310 was 5.4 times faster
• Significantly reduces overall system rollout costs and future operating costs
• Resolves storage device performance bottlenecks, releasing the existing performance of compute nodes
The Taiwan Chemical Engineering Research Team is one of the leading teams in Taiwan’s academic chemical engineering community. The team needs to make extensive use of quantum chemistry and materials science simulation software to research and test various chemical materials. As the complexity of the experiments increases, and due to different researchers needing to share the same IT equipment, it often takes a long time to produce experiment results. The team initially enhanced system computing performance by increasing compute nodes. However, after increasing to 15 compute nodes, there was still an obvious performance bottleneck.
AccelStor realized that the laboratory did not lack computing power; instead, the performance slowdown was in the back-end storage devices after all data had been processed. The traditional hard disk storage devices originally used by the laboratory were built using an SAS HDD RAID 0 configuration. The laboratory used 15 compute nodes and integrated these with the hard disk storage devices using an NFS protocol. AccelStor suggested replacing the back-end storage devices with a NeoSapphire P310, and adding a 10GbE network switch to increase the data exchange bandwidth (Figure 1). The P310 is also equipped with AccelStor’s exclusive FlexiRemap software technology, which targets the random write data that is likely to cause system storage performance bottlenecks, and improves performance acceleration operations.
Using the Gaussian 09 software frequently adopted by the research team, AccelStor applied the coupled cluster method to compare the overall computing time of a traditional hard disk array with the P310 All-Flash Array under the same structural model. The traditional hard disk array required 32 minutes 25 seconds, while the P310 only required 5 minutes 47 seconds, making it 5.4 times faster. Using the P310 to resolve storage bottlenecks in the original architecture enabled the compute node CPU to fully implement stoichiometry.
The AccelStor team also trialled and compared different quantities of compute nodes: When there was only one compute node, the traditional hard disk array’s performance was about the same as that of the P310. However, when compute nodes were increased and their quantity of data also increased, it was obvious that the P310 could still execute all operations within a certain period of time, while the traditional hard disk array required significantly more time (Figure 2).