Merge pull request #21 from cmu-db/final-presentation

Final presentation

Author: connortsui20

async-bpm

@@ -0,0 +1,21 @@
+{
+    "info": {
+        "title": "Eggstrain & Async Buffer Pool Manager",
+        "github": "https://github.com/cmu-db/15721-s24-ee1",
+        "description": "An asynchronous vectorized push-based execution and asynchronous buffer pool manager written in Rust.",
+        "students": [
+            {
+                "name": "Kyle Booker",
+                "url": "https://www.linkedin.com/in/ktbooker/"
+            },
+            {
+                "name": "Sarvesh Tandon",
+                "url": "https://www.linkedin.com/in/sarvesh-tandon/"
+            },
+            {
+                "name": "Connor Tsui",
+                "url": "https://www.linkedin.com/in/connortsui/"
+            }
+        ]
+    }
+}

proposal/15721-project.json

@@ -0,0 +1,289 @@
+---
+marp: true
+theme: default
+#class: invert # Remove this line for light mode
+paginate: true
+---
+
+# Eggstrain
+
+Vectorized Push-Based inspired Execution Engine
+Asynchronous Buffer Pool Manager
+
+<br>
+
+## **Authors: Connor, Sarvesh, Kyle**
+
+---
+
+# Original Proposed Goals
+
+-   75%: First 7 operators working + integration with other components
+-   100%: All operators listed above working
+-   125%: TPC-H benchmark working
+
+---
+
+# Design Goals
+
+-   Robustness
+-   Modularity
+-   Extensibility
+-   Forward Compatibility
+
+We made heavy use of `tokio` and `rayon` in our implementation.
+
+---
+
+# Refresher on Architecture
+
+![bg right:60% 100%](./images/architecture.drawio.svg)
+
+---
+
+# Refresher on operators
+
+-   `TableScan`
+-   `Filter`
+-   `Projection`
+-   `HashAggregation`
+-   `HashJoin` (`HashProbe` + `HashBuild`)
+-   `OrderBy`
+-   `TopN`
+
+---
+
+# Example Operator Workflow
+
+![bg right:70% 80%](./images/hashjoin.svg)
+
+---
+
+# Progress Towards Goals
+
+-   100%: All operators implemented, excluding `HashJoin`
+-   125%: TPC-H benchmark working for Q1
+
+---
+
+# Execution Engine Benchmarks
+
+Hardware:
+
+-   M1 Pro, 8 cores, 16GB RAM
+
+---
+
+![bg 90%](./images/csvreader.png)
+
+---
+
+# Correctness Testing and Code Quality Assessment
+
+We tested correctness by comparing our results to the results of the same queries run in DataFusion.
+
+Our code quality is high with respect to documentation, integration tests, and code review.
+
+However, we lack unit tests for each operator. We instead tested operators integrated inside of queries.
+
+---
+
+# Problem: In Memory?
+
+We found that we needed to spill data to disk to handle large queries.
+
+However, to take advantage of our asynchronous architecture, we needed to implement an **asynchronous buffer pool manager.**
+
+
+---
+
+# Recap: Buffer Pool Manager
+
+A buffer pool manager manages synchronizing data between volatile memory and persistent storage.
+
+*   In charge of bringing data from storage into memory in the form of pages
+*   In charge of synchronizing reads and writes to the memory-local page data
+*   In charge of writing data back out to disk so it is synchronized
+
+---
+
+# Traditional Buffer Pool Manager
+
+![bg right:50% 100%](images/traditional_bpm.png)
+
+Traditional BPMs will use a global hash table that maps page IDs to memory frames.
+
+*   Source: _LeanStore: In-Memory Data Management Beyond Main Memory (2018)_
+
+---
+
+# Recap: Blocking I/O
+
+Additionally, traditional buffer pool managers will use blocking reads and writes to send data between memory and persistent storage.
+
+Blocking I/O is heavily reliant on the Operating System.
+
+> The DBMS can almost always manage memory better than the OS
+
+*   Source: 15-445 Lecture 6 on Buffer Pools
+
+---
+
+# Recap: I/O System Calls
+
+What happens when we issue a `pread()` or `pwrite()` call?
+
+*   We stop what we're doing
+*   We transfer control to the kernel
+*   _We are blocked waiting for the kernel to finish and transfer control back_
+    *   _A read from disk is *probably* scheduled somewhere_
+    *   _Something gets copied into the kernel_
+    *   _The kernel copies that something into userspace_
+*   We come back and resume execution
+
+---
+
+# Blocking I/O for Buffer Pool Managers
+
+Blocking I/O is fine for most situations, but might be a bottleneck for a DBMS's Buffer Pool Manager.
+
+-   Typically optimizations are implemented to offset the cost of blocking:
+    -   Pre-fetching
+    -   Scan-sharing
+    -   Background writing
+    -   `O_DIRECT`
+
+---
+
+# Non-blocking I/O
+
+What if we could do I/O _without_ blocking? There exist a few ways to do this:
+
+-   `libaio`
+-   `io_uring`
+-   SPDK
+-   All of these allow for _asynchronous I/O_
+
+---
+
+# `io_uring`
+
+![bg right:50% 90%](images/linux_io.png)
+
+This Buffer Pool Manager is going to be built with asynchronous I/O using `io_uring`.
+
+*   Source: _What Modern NVMe Storage Can Do, And How To Exploit It... (2023)_
+
+---
+
+# Asynchronous I/O
+
+Asynchronous I/O really only works when the programs running on top of it implement _cooperative multitasking_.
+
+*   Normally, the kernel gets to decide what thread gets to run
+*   Cooperative multitasking allows the program to decide who gets to run
+*   Context switching between tasks is a _much more_ lightweight maneuver
+*   If one task is waiting for I/O, we can cheaply switch to a different task!
+
+---
+
+# Eggstrain
+
+The key thing here is that our Execution Engine `eggstrain` fully embraces asynchronous execution.
+
+*   Rust has first-class support for asynchronous programs
+*   Using `async` libraries is almost as simple as plug-and-play
+*   The `tokio` crate is an easy runtime to get set up
+*   We can easily create a buffer pool manager in the form of a Rust library crate
+
+---
+
+# Goals
+
+The goal of this system is to _fully exploit parallelism_.
+
+*   NVMe drives have gotten really, really fast
+*   Blocking I/O simply cannot match the full throughput of an NVMe drive
+*   They are _completely_ bottle-necked by today's software
+*   If we can fully exploit parallelism in software _and_ hardware...
+    * **We can actually get close to matching the speed of in-memory systems, _while using persistent storage_**
+
+---
+
+![bg 60%](images/modern_storage.png)
+
+---
+
+# Proposed Design
+
+The next slide has a proposed design for a fully asynchronous buffer pool manager. The full (somewhat incomplete) writeup can be found [here](https://github.com/Connortsui20/async-bpm).
+
+-   Heavily inspired by LeanStore
+    -   Eliminates the global page table and uses tagged pointers to data
+-   Even more inspired by this paper:
+    -   _What Modern NVMe Storage Can Do, And How To Exploit It: High-Performance I/O for High-Performance Storage Engines (2023)_
+        -   Gabriel Haas and Viktor Leis
+-   The goal is to _eliminate as many sources of global contention as possible_
+
+---
+
+![bg 90%](images/bpm_design.png)
+
+---
+
+# BPM Benchmarks
+
+Hardware:
+
+* Cray/Appro GB512X - 32 Threads Xeon E5-2670 @ 2.60GHz, 64 GiB DDR3 RAM, 1x 240GB SSD, Gigabit Ethernet, QLogic QDR Infiniband
+* We will benchmark against RocksDB as a buffer pool manager
+
+---
+
+![bg 90%](./images/zip1.1dist.png)
+
+---
+
+![bg 90%](./images/20w80r.png)
+
+<!--
+zipfian distribution, alpha = 1.01 -->
+
+---
+
+![bg 90%](./images/80w20r.png)
+
+---
+
+![bg 90%](./images/uniform20w80r.png)
+
+---
+
+![bg 90%](./images/uniform80w20r.png)
+
+---
+
+![bg 90%](./images/uniform5050.png)
+
+<!-- zipfian distribution, alpha = 1.01 -->
+
+<!-- ---
+
+![bg 90%](./images/zip1.1.png)
+
+zipfian distribution, alpha = 1.1 -->
+
+<!-- ---
+
+![bg 90%](./images/zip1.2.png)
+zipfian distribution, alpha = 1.2 -->
+
+---
+
+# Future Work
+
+-   Asynchronous BPM ergonomics and API
+-   Proper `io_uring` polling and batch evictions
+-   Shared user/kernel buffers and file descriptors (avoiding `memcpy`)
+-   Multiple NVMe SSD support (Software-implemented RAID 0)
+-   Optimistic hybrid latches