Designing an Energy-efficient Architecture for Geo Databases by Yichen Wei

The cost and efficiency of IP geolocation data services depends on careful system design. This talk covers how storage architecture, data compression, network optimization, memory layout, and CPU-level tuning all play a role. We also show how techniques like NUMA awareness, SIMD, zero-copy, and zero-serialization enable higher traffic with lower resource use.

1. A ScyllaDB Community
Designing an Energy-efficient
Architecture for Geo Databases
Yichen Wei
Engineer Manager

2. Yichen Wei
Engineer Manager
■HA Distributed Platform domains with
machine learning and GenAI
■Evangelist of HPC, HTAP, Rust and FP/CT
■WASM/WASI, FaaS

3. Geo service

4. Challenge

■Critical Service
●Low latency
●High availability
●Backward and forward compatibility
■Cost
●Data size
●Query performance

5. Geo service workflow

6. Why not existing local databases?
Why not using sqlite, duckdb or rocksdb?
■Different user cases
●Daily/Weekly full size dump
●No update operations
■Performance and cost concerns
●Read/Write amplification and memory footprint
●Long tail lookups

7. Solution

8. Geo service workflow

9. Optimization 1: Loading

10. Optimization 2: Parsing
■Encoding
●Dictionary encoding - Country, Region, City, Metro, Zip…
●Delta encoding - longitude and latitude

11. Optimization 2: Parsing
■Memory layout and alignment

12. Optimization 2: Parsing
■Memory layout and alignment

13. Optimization 3: Lookup
■

14. Optimization 3: Lookup
Chunk Size
■Graviton instance
●L1: 64K
●L2: 1M

15. Optimization 3: Lookup

16. Optimization 3: Lookup
SIMD
■
●

17. Optimization 3: Lookup
SIMD
■truth table
●16 bytes vector - 4 IPv4 value


IP1 IP2 IP3 IP4
T T T T
T T T F
T T F F
T F F F
F F F F

18. Optimization 3: Lookup
SIMD
■Does Golang have SIMD support?
●Yes and No

19. Optimization 3: Lookup
SIMD
■Does Golang have SIMD support?
●
●

20. Optimization 3: Lookup
SIMD
■Cgo is not Go!

21. Optimization 3: Lookup
SIMD
■https://github.com/alivanz/go-simd
■`linkname` directive to call C code, bypass CGO

22. Optimization 3: Lookup
SIMD

23. Optimization 3: Lookup
SIMD

24. Optimization 3: Lookup - One More Thing
premature optimization is the root of all evil

25. Optimization 3: Lookup - One More Thing

26. Optimization 3: Lookup - One More Thing

27. Optimization 3: Lookup - One More Thing

28. Optimization 3: Lookup - One More Thing

29. Optimization 3: Lookup - One More Thing
Understand compiler first before optimizations!

30. Looking Forward

■Compression & encoding
■Histogram Distribution
■Data prefetching
■Change Language

31. Takeaways

■No one knows your data better than you
■Don’t be afraid to build the infrastructure for your data
■Optimization is fun
●… and hard
■Golong doesn’t support SIMD well

32. Thank you! Let’s connect.
Yichen Wei
[email protected]