[Paper Review] Polynesia: Enabling Effective Hybrid Transactional/Analytical Databases with Specialized Hardware/Software Co-Design

Amirali Boroumand, Saugata Ghose+2 more|arXiv (Cornell University)|Mar 1, 2021

Distributed systems and fault tolerance64 references20 citations

TL;DR

Polynesia is a hardware-software co-designed HTAP database system that partitions workloads into isolated transactional and analytical processing islands, using specialized accelerators and processing-in-memory to reduce data movement and consistency overhead. It achieves 1.70X higher average transactional throughput and 3.74X higher analytical throughput than state-of-the-art systems, with 48% lower energy consumption.

ABSTRACT

An exponential growth in data volume, combined with increasing demand for real-time analysis (i.e., using the most recent data), has resulted in the emergence of database systems that concurrently support transactions and data analytics. These hybrid transactional and analytical processing (HTAP) database systems can support real-time data analysis without the high costs of synchronizing across separate single-purpose databases. Unfortunately, for many applications that perform a high rate of data updates, state-of-the-art HTAP systems incur significant drops in transactional (up to 74.6%) and/or analytical (up to 49.8%) throughput compared to performing only transactions or only analytics in isolation, due to (1) data movement between the CPU and memory, (2) data update propagation, and (3) consistency costs. We propose Polynesia, a hardware-software co-designed system for in-memory HTAP databases. Polynesia (1) divides the HTAP system into transactional and analytical processing islands, (2) implements custom algorithms and hardware to reduce the costs of update propagation and consistency, and (3) exploits processing-in-memory for the analytical islands to alleviate data movement. Our evaluation shows that Polynesia outperforms three state-of-the-art HTAP systems, with average transactional/analytical throughput improvements of 1.70X/3.74X, and reduces energy consumption by 48% over the prior lowest-energy system.

Motivation & Objective

• Address the performance degradation in state-of-the-art HTAP systems caused by data movement, update propagation, and consistency overhead.
• Overcome performance interference between transactional and analytical workloads due to resource contention.
• Achieve high throughput and low latency for both workloads simultaneously, matching isolated performance.
• Enable real-time analytics with data freshness and strong consistency guarantees.
• Design a system that meets all three ideal HTAP properties: workload-specific optimization, data freshness, and performance isolation.

Proposed method

• Partition the HTAP system into dedicated transactional and analytical processing islands, each with isolated data replicas, optimized execution engines, and tailored hardware resources.
• Implement custom hardware accelerators and algorithms for efficient update propagation between islands, minimizing CPU-to-memory data movement.
• Introduce a novel data consistency mechanism that ensures analytical queries see a consistent, up-to-date view of data without blocking transactional workloads.
• Leverage processing-in-memory (PIM) architectures in analytical islands to reduce off-chip memory accesses and accelerate analytical query execution.
• Design a runtime task scheduler that enables dynamic load balancing via task stealing across vault groups to improve utilization and throughput.
• Use a processor-centric interconnect topology to scale across multiple HMC memory stacks while maintaining low communication overhead.

Experimental results

Research questions

• RQ1 How can hardware-software co-design mitigate the performance degradation caused by data movement and consistency mechanisms in HTAP systems?
• RQ2 To what extent can workload isolation through dedicated processing islands eliminate performance interference between transactional and analytical workloads?
• RQ3 Can processing-in-memory (PIM) techniques significantly reduce data movement and energy consumption in analytical workloads within HTAP systems?
• RQ4 How does the system maintain data freshness and consistency while achieving high throughput for both workloads?
• RQ5 What is the scalability and energy efficiency of the co-designed system as data size and hardware scale increase?

Key findings

• Polynesia achieves an average of 1.70X higher transactional throughput and 3.74X higher analytical throughput compared to three state-of-the-art HTAP systems.
• The system reduces energy consumption by 48% compared to the lowest-energy prior system, primarily by minimizing off-chip memory accesses and using efficient PIM cores.
• Analytical throughput scales well with increasing data size, outperforming the multiple-instance baseline by up to 3.0X even as dataset size doubles across four HMC stacks.
• Transactional throughput degrades by only 8.8% when scaling from one to four HMC stacks in Polynesia, compared to a 54.4% drop in the multiple-instance baseline.
• Task stealing in the Hybrid-sched variant reduces remote memory access overhead and brings analytical throughput within 3.2% of the ideal remote-access baseline.
• The system maintains low update application latency while ensuring data consistency, demonstrating effective isolation between workloads without performance penalties.