Sabre: Hardware-Accelerated Snapshot Compression for Serverless MicroVMs
Highlights
-
Demonstrates a hardware-accelerated snapshot compression and prefetching system built around Intel's IAA for extremely quick low-latency decompression.
-
This paper feels like Intel Labs showing off how great their IAA is for this task. The way it's written, it feels like a "hardware vs software" page compression comparison more than anything. I think what they did is really cool and extremely novel on it's own, which makes it all the more frustrating that parts of the paper (i.e. the evaluations) aren't written better. I would have given it a weak accept.
Summary
To avoid cold start overheads, snapshotting allows part of the physical memory to be stored as a file. Later restoration of the snapshot from the file avoids the long cold start-up times. Pre-fetching can improve the effectiveness of snapshotting, though this is dependent on the size of the memory that needs to be restored. The authors introduce Sabre, a hardware-based snapshot page pre-fetcher and compressor that is shown to speed up memory restoration time from snapshots by up to 55% with a 4.5x compression factor.
Key Contributions
- Demonstrates a hardware-accelerated snapshot compression and prefetching system built around Intel's IAA for extremely quick low-latency decompression.
Strengths
-
The figures, graphs, and tables are very clean. The whole paper is put together well.
-
I think their background section does a very good job explaining the problem space as well as how their paper fits into it.
-
I like the amount of detail that went into the hardware and testbed side of the paper.
-
I really like both the presentation and comparison between single-chunk prefetching and scattered pre-fetching. It accurately demonstrates and explains the system trade-offs between running Sabre in one configuration versus the other.
Weaknesses / Questions
-
A considerable amount of time is spent talking about the IAA, which after reading this paper, I feel is doing 99% of the lifting. Saber seems like it's just along for the ride, despite being the title of the paper. However, the team that produced the IAA are also listed authors and disclosed the desire to provide a public spec of their hardware.
-
The evaluations in general honestly leave a bad taste in my mouth. They feel unneccessarily overly selective, or even cherry-picked. (e.g. "We do not show results for LZ4 as its compression ratios are low"). For a general pre-fetcher and compressor, I expect to see more system centered datasets but 11 out of 13 of them are compression centered, which might not represent realistic workloads.
-
The evaluations in general can be summarized as "this is how our hardware vastly beats out any software method out there" (e.g. Figure 9). That's fine, and I think the paper demonstrates how great their IAA is, but after accounting for hardware acceleration, the benefit that Sabre provides feels modest at least.
-
As a follow up example to evaluations, they omitted the effect of page faults on throughput for hardware vs. software. Inclusion would have given me a comprehensive picture of the worst case.
-
Nit: The black text on maroon and dark grey background is difficult to read.
Related Work
-
MicroVMs
-
VM Snapshotting
-
VM Pre-fetching