#Intro

Jacob Stanley
Icicle (?)
Work at Ambiata
- Use simple models to predict the future
- Costs more to run code than to develop
- Correctness more important than performance
- Optimized DSL

Icicle

Query language for time-series data
- latest 5 in mean value <- Example icicle query
Three optimizations!

Schema Specialization

Code is I/O Bound -> its 1996?
- 1800 MB/s sequential reads on macbook
- Fastest ssd is 3300 MB/s sequential reads
- You are not I/O bound - you're just slow
Parsing performance matters
- Avoid reading one byte at a time
- (Read multiple at a time)

The perfect parser

Exploit what we know about "THe Scheme"
- ASD|lang|{"sh: true"}|2016-01-01\n
Memchar
- Data.ByteString.elemIndex in haskell
The scheme contains well known information
- Strings are just lists of 64 bit integers
Buffers always padded with 8 bytes
Read multiple bytes at a time

Partial Evaluation

icicle is pure and total
Beta reduction
- Process of applying lambda to an expression
Constant folding
- Evaluation of constants and operations at compile time
Icicle source -> Icicle core
- Things are compressed / inlined
PE Bottom up
Beta reduction & Constant folding done at the same time
- Lambdas are removed
- necessary to compile to C

Melting Down Data

Records melt down to individual values
Arrays melt down to individual arrays
Options melt down to individual values + is some bool
Unpack - extract melted value (?)
Pack - Smash melted values back together

Optimization Opportunities

Eliminating dead code
- Only using high values
Eliminating compound types
- Can be easily / quickly compiled to C