This topic describes the Distributed Memory Pool (DMP) technology in PolarDB. With DMP, PolarDB implements a three-layer architecture that decouples compute, memory, and storage. This technology significantly improves read and write performance for I/O-intensive payloads using a multi-node distributed high-performance cache and a high-speed, one-sided Remote Direct Memory Access (RDMA) network. You can dynamically enable or disable DMP in seconds without any impact on your services and without interrupting or blocking cluster operations.
This feature is in canary release. To use this feature or if you have any questions, submit a ticket.
Scope
Edition: Enterprise Edition.
How it works
A database buffer pool is an area in memory used to cache disk data. When data is requested, the system first attempts to read it from the buffer pool. If the data is not in the buffer pool, it must be retrieved from the underlying storage. This I/O operation increases read latency, which includes both disk latency and network latency for remote storage. This directly affects database performance.
Due to the resource limitations of a single host, PolarDB clusters are offered with a fixed ratio of CPU to memory. This model presents three challenges:
Inflexibility for different business scenarios: For example, memory-intensive services require large amounts of memory, while compute-intensive services require more CPU cores. If compute and memory were decoupled, services could be configured more flexibly to reduce costs and improve efficiency.
Limited memory capacity: The maximum memory of a cluster is limited by the maximum memory of a single host. For example, most current cluster specifications have a maximum memory of 512 GB.
Service impact during scaling: Scaling resources requires migration, which can cause transient disconnections to your services. This means that scaling for large-scale business events must often be planned days in advance, making it difficult to handle sudden traffic bursts.
PolarDB DMP uses a one-sided RDMA high-speed network to combine the memory of multiple hosts into a large-capacity, low-latency, and scalable distributed shared memory pool. After DMP is attached, PolarDB compute nodes automatically handle I/O requests by searching for data in the following order: local cache, local memory, distributed memory, and then distributed storage. This process is transparent to your services and requires no changes to your applications. Resources can be scaled dynamically after attachment.
The attachment and scaling processes do not impact your services.
With DMP, the maximum memory of a single PolarDB cluster can reach 10 TB. The 16 KB random read I/O latency is as low as 10 us, which is approximately one-tenth of the storage I/O latency. This significantly improves performance in I/O-bound scenarios, with up to a 6.5× increase in TPC-H performance.
Core advantages
Improved performance for I/O-bound scenarios: With one-sided RDMA, DMP reduces the 16 KB random read I/O latency to as low as 10 us. This is 10 times lower than the I/O read latency from storage.
Breakthrough memory capacity: The maximum memory for a single node can reach 10 TB.
Seamless elasticity in seconds: Attaching DMP takes less than 1 second and does not impact your services.
Flexible configuration: You can configure DMP at the cluster level or for the primary node and specific read-only nodes individually to improve read and write performance.