In a recent article for The Register, Chris Mellor discussed PowerMax and some of our plans for 2019. In summary, we designed PowerMax specifically because we could see Storage Class Memory (SCM) as the coming media for the most important (and I/O intensive) data. We designed PowerMax to use all NVMe media, so that SCM would be just another type of capacity we could exploit. We designed the encryption so that SCM would be protected just like our other media. The Machine Intelligence classifies workloads, so that the most important data can get the benefit of this new (and expensive) media without any additional effort from customers using PowerMax arrays – it will just work. As we noted at Dell Technologies World last year, we plan to deliver SCM for all PowerMax systems this year. And to complete the NVMe picture, we plan to ship NVMe over Fibre Channel fabrics this year as well, giving us NVMe from the host to cache to the drives, providing customers with even lower latency I/O.
When we launched PowerMax at Dell Technologies World last year, we talked about the change to NVMe and our focus on latency. One of the slides in my presentation covering PowerMax was focused on how we make that happen.
The caching of data for low response times is historical. This has been a part of VMAX3 and VMAX All Flash systems. The system change to NVMe for drive attachment provides the PowerMax systems with significant benefits in latency, IOPS, and RAS. Since the NVMe drives are dual ported, each director has dual paths to all media that are always active. It also allows both directors in a brick to share access to all media at the same time, driving up performance and driving down latency. Multiple I/O queues allow for the separation of large and small block reads and writes.
To be clear, the data path from the CPUs to the drives is all switched PCIe, with NVMe as the access protocol. This ensures that we have the lowest latency for storage media access. The design also provides D@RE encryption support in the backend I/O modules, so that all data can be encrypted at line rates without regard for the type of drives in use[1].
Data placement is managed by Machine Intelligence. The coldest data, or data classified at the lowest service levels, will be reduced and stored on the slowest media. By understanding the nature of the workloads on the array, PowerMaxOS can reasonably predict the workload patterns that will occur in the near future.
PowerMax provides hardware data reduction. The design provides optimized response times and capacity management, delivering full bandwidth even for heavily compressed data. Based on the intelligent data placement, the most active data will be stored on SCM media, so it will not be reduced (and not have the ~50us of read latency added for decompressing data on reads).
PowerMax service levels ensure that the system is aware of the relative importance of data on different storage groups. Data classified as Bronze will never get SCM capacity, while data classified as Diamond will be given the most access to SCM.
Moving from SAS to NVMe drive connections has lowered the media latency for PowerMax. The additional queues with NVMe allow the drives to sustain dramatically higher workloads while maintaining lower latencies.
Currently, all NVMe flash drives used in PowerMax systems offer the same performance/latency. As we introduce Storage Class Memory media later this year, the system will automatically make use of that capacity to ensure that the most active and important data is stored on this new class of media. After all, SCM is going to be significantly more expensive than the NAND media for the near-term, so we want to help customers get the most value from it.
To further reduce latency, we will ship NVMe over Fibre Channel fabrics this year, giving us NVMe from the host to cache to the drives. Building the full solution requires HBAs that support NVMe over FC, a fabric of gen 5/6 switches, and array support. We are doing our usual deep testing of the entire stack, and will ship with a complete support matrix that customers can count on for production operations.
Since most of the host multi-path drivers are based on SCSI intercepts, they have to be rewritten. To ensure that we have the most reliable infrastructure options for customers, a version of PowerPath will also be available to manage the new NVMe multi-pathing needs.
Changing from SCSI over Fibre to NVMe cuts out the heavy SCSI protocol stack, reducing latencies from the host to the array by about 50us, continuing our drive to minimize the time for customer applications to get the data they are seeking. Which gets added to the lower latency from SCM drives. Which gets added to service level management and hardware-based data reduction. Which gets added to the intelligent placement of data and optimized data reduction. Which is added to the NVMe backend for all storage media. Tied together with intelligent data management and a huge, robust cache. All leading to PowerMax as the most powerful array you can buy.
For more detailed background on PowerMax in video, you can find my technical presentation to Storage Field Day #16 here.
[1] Some competing arrays have announced plans to use SCM as a caching tier. Since SCM is non-volatile, that would mean encrypting data into and out of cache, which the systems are not currently doing (so no D@RE with SCM). And many CPUs do not allow hot swap of memory (like SCM DIMMs).
