The 1000x Spreadsheet: Performance from first principles

Follow Causal's journey of scaling their spreadsheet engine from thousands to billions of cells. Profilers are helpful but first-principle thinking is required to break out of iterative improvements and make order of magnitude improvements.

Transcript

1. The 1000x Spreadsheet Performance from first principles 
 @angadn, Causal
2. Simple phone-app business model
8. Categories add cells exponentially 💣 10 variables x 100 months x 10 products x 10 countries 
 
 x 10 languages 1,000 cells 10,000 cells 100,000 cells 1,000,000 cells
12. Atwood’s Law 👨⚖ - Single Threaded - Event loop woes - Workers and data serialization - Interpreted environment - Preallocation discouraged - Everything is a hash-map
13. Golang or go home 💯
15. Profiling for fun and profit - go tool pprof - Preallocation, String concatenations - sync.Pool
16. Causal profiling with coz
17. - Initialise what you can upfront. - Channels are for productivity, not performance. Use errgroup. - Anything that needs safety can go into a Context. - Run go test --race in CI! Practice safe concurrency 🫃
20. How much faster can we go?
21. Establishing a baseline with Napkin Math™ Rockets need metals. Software needs CPU, memory, and I/O. Simon Eskildsen sirupsen.com/napkin
22. “I believe that first-principle thinking is required to break out of iterative improvement and make order of magnitude improvements.”
23. Napkin Math 🧠 - 1 variable with 1 Billion cells = 8 GB - Adding 2 variables = (2 + 1) * 8 GB = 24 GB - With SIMD and threading: 30 ms / 1 GB - 24 GB * 30 ms ≅ 700 ms
24. 👎 - If an array is filled up 50% or so, is it really sparse? - The real cost of maps will never show up on profiles. - Flame charts don’t show CPU cache misses. Maps in Golang are really fast 🤷 👍 - Golang hashes things really fast. Especially integer-keys.
25. ELI5: CPU Caches 👶 - Temporal Locality = LRU - Spatial Locality = Cache lines - Typically, each cache line is 64 bytes
26. CPUs ❤ sequential access
27. Anti Moore’s Law, DDR5 vs. 4: 2x throughput, but +12% latency.
28. How do we prove our hypothesis? 🧐 - The real world has too much variance (esp. on public clouds like GCP/ AWS). - Simpler POC → Faster iteration → More experimentation. - The simplest Causal model is adding up two variables with millions of cells.
29. 6 seconds → 400 milliseconds 🏎 💨
30. Structs-of-arrays, Out-of-bands… - Smaller structs - When storing references: Pointers are 64-bit, integers are 32-bit. - Structs should reflect locality in data-access.
32. Switching to Intel Ice Lake on GCP sped us up by 30-50%.
33. The road ahead 🚏 - Vectorised Arithmetic + SIMD - Time-series Compression - GPUs - Distributed Monte Carlo simulations - Sparse data structures …
34. causal.app/careers @causalhq @angadn