Measurements
Prefixes:
- mili: $m$, 0.001,3,
- micro: $\mu$, 0.000 001, 6,
- nano: $n$, 9,
- pico: $p$, 12,
- femto: $f$, 15,
- atto: $a$, 18,
- zepto: $z$, 21,
- yocto: $y$, 24
Speedup
- Amdahl’s law: SpeedUp= $\frac{1}{1-p+p/s}$. Constant problem size.
- Gustafson’s law: SpeedUp=$1-p+sp$. Constant computing power.
Tuning
- scale out
- scale up: memory, disk, CPU, network … an easy but last resort
- code:
- look for large loops
- avoid exception catching
- avoid polymorphism
- avoid virtual function
- go low level if needed
- size of chunks: make the size smaller for liquidity, but not too much for latency issues.
- storage format: syntax VS binary format
- network usage:
- keep shuffling to a minimum
- push down projection and selection as close as possible to the source.
- architecture
Tail latency
Causes
sharing resources, queues, garbage collection, energy management…
Improvements
- keep the low-level queues short
- break splits and queries further down
- manage background activities
- latency-induced probation: we shut down a machine that is too slow.
- completely drop the last 0.01% that takes too long…
- other measures
Hedge requests
basic idea: duplicate
task duplicates, first done wins.
Deferred hedge request
when 95% percentile reached, launch duplicate tasks.
Tied request
basic idea cancel request. A task is duplicated and waiting in several queues. Cancel other requests when one of them starts.
