Leading hyperconverged infrastructure (HCI) provider SmartX announced its latest high-performance, low-latency HCI solution and SmartX Halo P series appliance built with SmartX’s core software SMTX OS and Intel® Optane™ persistent memory (PMEM). For the first time in the industry, SmartX has applied PMEM with App Direct mode as a key component in HCI solution. At the same time, with end-to-end optimizations across computing virtualization, storage network as well as storage medium, the solution obtains industry-leading performance as well as low-latency server virtualization storage based on limited computing resources. During the test, only 3-node SmartX Halo P series appliance is applied to reach 1.2 million IOPS, 25 GB bandwidth, while the latency at the application side keeps below 100 microseconds.

1.jpg

With outstanding stability and leading features, the HCI solutions from SmartX have been widely applied to the production environments in banking, securities, insurance, funding as well as other financial verticals. However, some clients are hesitated to replace their bare metal servers and all-flash array with HCI solution, over the concerns of the performance and latency of computing virtualization, networking and storage layers. To solve this problem, SmartX started a project with the code name “Sailfish” (the fastest fish in the ocean) since 2019, with App Direct mode for cache acceleration and fully optimized SmartX HCI core software SMTX OS.

Intel Optane PMEM adds a new storage tier between memory and SSDs, which has memory-like high performance with ultra-low latency, high durability and reliability, data persistency, as well as byte-addressable. Configured in App Direct mode, the applications are able to access the data in persistent memory of Intel Optane PMEM.

To fully leverage the capabilities of PMEM, SmartX has made end-to-end optimization across computing virtualization, storage network, and storage medium. Due to the specific architecture of HCI, the performance enhancement can only be achieved with limited computing resources, which has brought with many challenges for the Sailfish project. To counter those difficulties, SmartX has made improvements from the following perspectives.

  1. At the storage layer, SmartX fully leverages the low latency and persistent memory of PMEM App Direct mode to save the most frequently accessed data in the cluster. In order to ensure the reliability and high availability of the data, the data cached in PMEM will also be replicated to the other nodes. Not only that, SMTX OS takes full advantage of the byte-addressable feature of persistent memory to redesign the journal, effectively solving the problem of write amplification. Through the DMA method, the data copy tasks between the memory and the persistent memory are transferred to the hardware, greatly improving the efficiency of memory copy, so as to enhance the storage performance, without generating additional CPU resource requests.
  2. At the computing virtualization layer, the storage virtualization is devolved from the virtual machine to the storage software stack, through the SMTX ELF boost mode to avoid additional performance overhead caused by IO requests passing through the virtual machines. At the same time, the memory is shared by the VM and the storage system to avoid memory replication on the IO path.
  3. At the storage network layer, RDMA over Converged Ethernet (RoCE) is used to accelerate network IO requests. The implementation of the network protocol is transferred to the network card. Hardware acceleration function of the network card is leveraged to reduce the distributed storage performance overhead during network communications.

With all the above optimizations, the virtual machine’s IO performance has been enhanced by 320%, while latency reduced by more than 85%.

3.jpg

So far, the technologies used in Sailfish project have been integrated into the SmartX Halo P series appliance. SmartX Halo P series appliance is capable of mission-critical applications such as transaction databases and machine learnings with demanding requests on IO. In addition, with higher performance, SmartX Halo P series appliance also increases the density of virtual machines to simplify IT deployment and reduces TCO. Currently, the appliance has been tested and implemented in the financial industry already.

Compared with a mainstream mid-to-high-end all-flash array in the market, the 8KB random read performance of the 3-node SmartX Halo P series appliance reaches 1.2 million IOPS, which is 1.4 times that of the AFA. The sequential read bandwidth reaches 25 GB, 2.1 times of AFA. The application side latency is only 100 microseconds, while the latency of the AFA at controller end reaches 500 microseconds already. Moreover, the Halo P series appliance keeps the simplicity and resilience of HCI architecture, as well as the highly competitive TCO compared with mid-to-high level AFAs.

“SmartX has been committed to building the most excellent HCI infrastructure.” According to Kyle Zhang, Co-founder and CTO of SmartX,” In the Sailfish project, we provide the best practice in adopting PMEM in HCI solutions. From the test results, we have seen that the introduction of new storage technologies can greatly improve the performance of HCI system and address the real-workload challenges for critical applications. In the future, SmartX will collaborate with Intel and other leading industry leaders to introduce more advanced technologies to lead the next-level innovations in HCI.”