Pei Chen
I am a senior software engineer in Huawei Cloud Computing Technology Ltd, Beijing. I am working on improving metadata performance of file and object storage systems in cloud computing.
Prior to industry, I earned my MPhil in Computer Science from Tsinghua University(THU) in 2022, co-advised by Professor Jiwu Shu and Professor Youyou Lu in Storage Research Group.
Previously, I obtained my B.S. in Computer Science at Central South University(CSU), advised by Professor Weiping Wang.
Please find my CV here.
GitHub Twitter Linkedin chenpei.2018AT tsinghua.org.cn
Research Interest
I work in the area of System for Cloud Computing and database systems. My research spans distributed systems, and storage systems on emerging hardware (e.g., Non-Volatile Memory, RDMA).
In particular, I am interested in building distributed systems, key-value stores and transactional systems leveraging emerging devices (e.g., RDMA, SmartNICs). Moreover, I am also passionate about hardware memory disaggregation, enabling independent scaling of memory and compute.
What's
New
[Jan 2022] I earned my MPhil in Computer Science from Tsinghua University(THU).
[Jun 2018] I obtained my B.S. in Computer Science at Central South University(CSU).
Research
Experience
Below are my research experience and industry projects.
-
SNlog: A SmartNIC-driven shared log
Preprint.
Pei Chen, Youyou Lu, Qing Wang, Junru Li, Jiwu Shu. What: A disaggregated shared log system that offloads log operations to emerging SmartNICs.
Benefits: Better throughput compared to existing log systems, without consuming significant computational resources on the server side. and fully utilizing the bandwidth of raw SSD devices.
How: Leverage the ARM-based Mellanox BlueField-2 SmartNIC, direct read and write operations to flash storage devices are performed. Propose techniques include delayed persistence log management, efficient concurrency control mechanisms and a replica protocol separates metadata and data. A two-level mapping mechanism enhances log read performance and addresses the limited memory challenge of SmartNICs. RDMA is utilized for data transmission, minimizing message copying. -
Efficient and Consistent NVMM Cache for SSD-based File System
TC 2018.
Youmin Chen, Youyou Lu, Pei Chen, Jiwu Shu. What: Use persistent memory as SSD cache.
Motivation: Using persistent memory as SSD cache can reduce the data persistence overhead, but the existing cache construction method does not consider the workload characteristics and cannot give full play to the hardware performance.
How: Build adaptive cache management (AFCM) combining pages and cache lines, which reduces the data write amplification of coarse-grained cache management (SCCM), and eliminates the cache index overhead caused by the fine-grained cache management (SFCM) with the granularity of the cache line. Designed copy-on-write (TCOW) mechanism to ensure the crash consistency of NVMM cache space and file system mirroring.
Benefits: The performance of AFCM is 83% higher than that of SCCM, and the write operation to SSD is reduced by 63%. -
A global address space management method for distributed persistent memory
Patent No: CN111241011A.
Jiwu Shu, Youmin Chen, Qing Wang, Pei Chen, Youyou Lu What: Adress space management for distributed persistent memory.
Motivation: It is essential to build a distributed persistent memory allocator with high availability, high performance and crash consistency, but the existing persistent memory allocator cannot satisfy the above characteristics at the same time.
How: Use a centralized monitor to manage the node and address mapping information in the cluster. Use the raft protocol to ensure the high availability of the monitor. Use the primary backup protocol to ensure the high availability of the data. Build a thread-isolated address mapping table and batch processing strategy to improve access to metadata server performance. Use the state machine mechanism inside the local allocator, and use the two-phase commit method to ensure the atomicity of address allocation. Design a lock-free data structure for address allocation -
A memory communication method and device based on RDMA
Patent No: CN111858418A.
Youyou Lu, Jiwu Shu, Youmin Chen, Pei Chen, Jun Xu, Peng Lin. What: Deliver scalable performance on RC RDMA.
Motivation: RDMA suffers performance collapse when transferring data to an increasing number of clients on reliable connection because of resource contention in the CPU cache.
How: Deliver scalable performance based on RC RDMA. It achieves so by grouping and rating the network connections, so as to balance the saturation and thrashing of the cache space. Introduced a background thread pool for efficient data storage and enhanced RDMA with strategic scheduling. Constructed a QP-isolated RDMA communication method while utilizing GPUDirect for inter-node communication, reducing message copying overhead. Developed a background thread pool for parallel data persistence and improved RDMA scalability through hardware resource reuse and efficient scheduling. Achieved a write bandwidth 18\% higher than the baseline system.
-
Introduce vector engine of ClickHouse into MySQL
Research Intern. Jan 2022 - April 2022
ByteDance Corp · Beijing What: Designed and implemented a c++ source code MysqlExecutor by imitating ScanExecutor of ClickHouse to support ByteNDB storage. Developed a presentation of the new schema and drafted a list of C++ classes that needs to be modified to support ByteNDB storage.
-
Optimization of Local Storage Read Performance and Reliability for Cloud Storage Metadata
Senior Software Engineer. May 2022 - Sep 2022
Huawei Cloud Corp · Beijing What: Improving point-lookup using data block hash index. . Do some basic consistency checks to LSM-tree
Benefits: The CPU utilization of one of the main functions in the point lookup code path, DataBlockIter::Seek(), seek is reduced by 21.8% and the overall throughput is increased by 10% at an overhead of 4.6% more space. . Do some basic consistency checks to LSM-tree
-
Scalability and Reliability Optimization of Cloud Storage Metadata Based on MongoDB
Senior Software Engineer. Oct 2022 - June 2023
Huawei Cloud Corp · Beijing What:1) Reliable routing point lookup strategy based on MongoDB, reducing cloud storage metadata access failure duration from 30 minutes to 30 seconds. 2) MongoDB-based route scalability optimization based on mongoDB, eliminating invalid routes, optimizing lock contention, and reducing access latency from 10 seconds to 0.6 milliseconds, a reduction of 2 orders of magnitude 3) In-place assign based on MongoDB chunk, reducing blast radius, improving chunk takeover failure SLA from 96% to 99.99%.
-
Building a Columnar Analytical Engine Based on Object Storage LifeCycle Service
Senior Software Engineer. July 2023 - Sep 2023
Huawei Cloud Corp · Beijing What: Designed and implemented a Java source code MergeEngine focusing on object metadata and version management, which can process tasks based on job style. It Eliminated the secondary index for Object Life cycle, in order to improve object storage scalability. Meanwhile, it reduced the cycle of AP update operations from 1 day to 15 minutes. Developed a presentation of the new architecture and drafted a list of Java features that were not yet properly supported by the existing source codes.
-
Operator Parallelization Support Based on SMP
Senior Software Engineer. Oct 2023 - Present
Huawei Corp · Beijing What: Designed and implemented SMP-based operator parallel processing, including insert into select, merge into, update, and delete operators. The speed-up can reach up to 0.8 when the OPS is improved by more than 5 times; implemented SMP-based parallel shared hash join through decreasing the data copy and memory consumption. The OPS increase can reach about 2 times.
Education
Experience
By reserved by time order
Tsinghua University
2018.09 - 2022.01 Advisor: Prof. Jiwu Shu & Prof. Youyou LuI had great experience working with Prof. Jiwu Shu, and Prof. Youyou Lu. This experience strengthens my problem-solving and communication skills and broadens my research horizons.
Central South University
2014.09-2018.07 Advisor: Prof. Weiping WangI am lucky to follow Weiping who encourage me to go for professional competence.
Services
[2022 - 2023] Lecturer, Tsinghua University Student Career Development Association Student Tutor Group.
[2021 - present] Member, Morning Jogging Team of Tsinghua University.
[2021 - 2023] Member, Bodybuilding Team of Tsinghua University.
[2021] Guest speaker, Tsinghua University Computer Science Department "Future of Computing" Phd and Master's Forum.
Awards
🏆[2018] Outstanding Undergraduate Thesis Award(Top 2%)(TODO:加个pdf链接) of CSU.
🏆[2018] Outstanding Graduate(Top 0.1%) of Hunan Province & CSU.
🏆[2017] Second-class Scholarship(Top 5%) of CSU.
🏆[2017] Excellent Student(Top 1%) CSU.
🏆[2016] National Scholarship(Top 0.2% Nationwide), China.
🏆[2016] First-class Scholarship(Top 1%) of CSU.
🏆[2016] Qu Yuan Scholarship(Top 0.1%) of CSU.
🏆[2016] Excellent Student Cadre(Top 2%) CSU.
🏆[2016] Honorable Mention of the International Mathematical Contest in Modeling(MCM).
🏆[2016] Second Prize of University Students' Energy Conservation and Emission Reduction Competition of China(Top 5% Nationwide).
🏆[2016] Second Prize of Information Security Competition of CSU.
🏆[2016] Third Prize of Programming Competition of CSU.
🏆[2016] First Prize of Transportation Science and Technology Competition of CSU.
🏆[2016] First Prize of Energy Saving and Emission Reduction Competition of CSU.
🏆[2015] National Encouragement Scholarship(Top 5%), China.
🏆[2015] Second-class Scholarship(Top 5%) of CSU.
🏆[2015] Chang Wei Scholarship(Top 1%) of CSU.
🏆[2015] Excellent Student Cadre(Top 2%) CSU.
🏆[2015] Third Prize of the Chinese Mathematical Contest in Modeling(MCM).
Miscellaneous
I like running long distances.
I am learning to ski.