DmRPC: Disaggregated Memory-aware Datacenter RPC for Data-intensive Applications

DmRPC is a fast RPC lib for Data-intensive Applications and Disaggregated Memory, which is the first DM-aware datacenter RPC for data-intensive datacenter applications, our ICDE 2024 paper describes the system in detail.

Some highlights:

• Pass by reference: use pass by reference instead of pass by value.
• Multiple supported disaggregate memory: support Network(RoCE/Ethernet) and CXL.
• Easy to program: use copy-on-write mechanism to provide programming simplicity.
• Much faster than Ray and Spark: DmRPC designed for fine-grained microservices. In our micro-benchmark, DmRPC achieves up to 21.6 $\times$ / 5.6 $\times$ higher throughput compore with Ray/Spark.
• High performance: get 3.1 $\times$ higher throughput and 2.5 $\times$ lower average latency than the baseline on DeathStarBench test suites, and get great performance for synthetic 7-tier microservices.

Architecture

DmRPC-net Architecture

System requirements

• NICs: 100GbE+ NICs are needed for good performance. DmRPC works best with Mellanox Ethernet(Connect-Serial) NICs. Any DPDK-capable NICs also work well.
• HugePage: At least 2048 huge pages on specific NUMA node

Directory Structure:

├── cn  (client node)
│   ├── app
│   │   ├── address_translate   (address translate speed test)
│   │   ├── bandwidth_read_test  (read bandwidth test)
│   │   ├── bandwidth_write_test (write bandwidth test)
│   │   ├── fork_speed (create reference speed test)
│   │   ├── fork_test (cow ratio speed test)
│   │   ├── img_transcode ( synthetic 7-tier microservices test)
│   │   ├── latency_test 
│   │   ├── simple_test
│   │   └── social_network (death star bench test suite)
│   └── rmem_ulib (client DmRPC user lib)
│       └── impl
├── extra
│   ├── ray
│   └── spark
├── include
├── mn (DmRPC memory node )
├── scripts (Test scripts)
├── third_party
└── util

Dependency:

Package	Version
mongoc	1.7
protobuf	3.21
dpdk	21.05
g++	>7.5.0

QuickStart

• Build and run memory node:

  • install dependency packages
  • mkdir build; cd build && cmake ..; make -j
  • run sudo ./rmem_mn

• Run different tests

  • replace folder name in /cn/app/build_app
  • cd build && cmake ..; make -j

• Run ray/spark

  • to /extra folder and see ray/spark test code

How to deploy different machine's IPs and ports?

• create a file named app_process_file in /cn/app folder
• write your IP and port in this file
• run you program with --client_index and --server_index with specific index in this file

ThirdParty

project	Version
asio	latest
eRPC	latest
atomic_queue	latest
parallel_hashmap	latest
HdrHistogram_c	latest
json	latest

Getting help

Working in the process...

Contact

email at [email protected]

Cite

@inproceedings{zhang2024dmrpc,
  title={DmRPC: Disaggregated Memory-aware Datacenter RPC for Data-intensive Applications},
  author={Zhang, Jie and Chen, Xuzheng and Zhang, Yin and Wang, Zeke},
  booktitle={2024 IEEE 40th International Conference on Data Engineering (ICDE)},
  pages={3796--3809},
  year={2024},
  organization={IEEE}
}