work in progress slides
This commit is contained in:
parent
32c4468700
commit
e8c8730157
|
@ -0,0 +1,37 @@
|
||||||
|
---
|
||||||
|
template: templates/cse662_2019_slides.erb
|
||||||
|
title: Thread Pinning Tricks
|
||||||
|
date: October 28
|
||||||
|
---
|
||||||
|
|
||||||
|
<!--
|
||||||
|
- Crescando Overview
|
||||||
|
- Goals:
|
||||||
|
- Low latency vs Predictable Latency
|
||||||
|
- Workload: OLTP-style; Simple, but high throughput queries and updates
|
||||||
|
- Challenges:
|
||||||
|
- Inter-query interference (Locking is a no-go)
|
||||||
|
- Caching: High variability in data resources being accessed
|
||||||
|
- Observation:
|
||||||
|
- Lots of overlap between queries in workload
|
||||||
|
|
||||||
|
- Miniaturize:
|
||||||
|
- Assign a CPU thread to a chunk of data
|
||||||
|
- Maximize throughput on this one chunk
|
||||||
|
- Branching a problem
|
||||||
|
- Cache lines a problem
|
||||||
|
|
||||||
|
- Coalescing Operations
|
||||||
|
- Classic Scan: Loop over Individual Operations
|
||||||
|
- Elevator Scan:
|
||||||
|
- Loop over Data
|
||||||
|
- Stream in each chunk of data to all operators in parallel
|
||||||
|
- Allow operators to attach in between any chunks (detach after 1 full scan)
|
||||||
|
- Discuss:
|
||||||
|
- Why is this ok / better?
|
||||||
|
- When does this fail?
|
||||||
|
-
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
-->
|
Loading…
Reference in a new issue