Главная
Study mode:
on
1
Intro
2
In-Memory Applications
3
Memory Is Inelastic
4
Trending Solution: Far Memory
5
Existing Far-Memory Systems Perform Poorly
6
Why Do Existing Systems Waste Performance?
7
Challenge 1: Semantic Gap
8
Challenge 2: High Kernel Overheads
9
Design Space
10
Remoteable Data Structure Library
11
Userspace Runtime
12
Pauseless Evacuator
13
Remote Agent
14
Sample Code
15
Implementation
16
Evaluation
17
Performance on Different Compute Intensities
18
NYC Taxi Analysis (C++ DataFrame)
19
Other Experiments
20
Related Work
21
Conclusion
Description:
Explore a groundbreaking approach to memory management in datacenter servers through this 20-minute conference talk from OSDI '20. Dive into the concept of Application-Integrated Far Memory (AIFM), a high-performance solution that addresses memory scarcity and underutilization in modern datacenters. Learn how AIFM enables applications to access remote memory resources with local RAM-like latency, avoiding read and write amplification issues common in paging-based systems. Discover the innovative API that allows developers to create remoteable, hybrid near/far memory data structures while maintaining transparency and ease of use. Examine the key insights behind AIFM's efficiency, including the exposure of application-level semantics to a high-performance runtime. Explore real-world applications of AIFM, including a web application frontend, NYC taxi data analytics, a memcached-like key-value cache, and Snappy compression. Compare AIFM's performance against state-of-the-art kernel-integrated, paging-based far memory systems, and understand its potential to revolutionize memory utilization in datacenter environments. Read more

AIFM - High-Performance, Application-Integrated Far Memory

USENIX
Add to list
0:00 / 0:00