Target Duration: 60–90 seconds (~180–220 spoken words)
Goal: Deliver a clear, plain-English overview of why the project was built, how the mechanism works, and key results, avoiding heavy jargon while dropping strategic follow-up hooks for the interviewer.
"There is a research paper called CRISP: Critical Slice Prefetching, published at ASPLOS 2022 by Heiner Litz, Grant Ayers, and Parthasarathy Ranganathan. In my project, I implemented a simplified, CRISP-inspired mechanism in ChampSim to understand its core ideas and evaluate their impact on processor performance.
The paper addresses a common problem in out-of-order processors: long-latency memory accesses can stall dependent instructions and limit instruction-level parallelism. Its key idea is to identify critical loads and prioritize the instructions in their dependency chains so that load addresses can be generated earlier, helping hide memory latency.
My implementation has three main components.
First, I added dependency tracking at dispatch. I track the most recent writer of each register and record predecessor instruction IDs to reconstruct register-dependency chains.
Second, I implemented runtime slow-load detection and backward slice construction. At retirement, I measure memory-instruction latency and compare it against an adaptive threshold. When a load exceeds the threshold, I use breadth-first search to trace backward through its recorded dependencies and mark the resulting instructions as critical.
Third, I modified the execution scheduling policy to prioritize ready critical instructions over non-critical ones, while giving the ROB-head instruction an early execution opportunity to help maintain forward progress.
I evaluated my implementation on five benchmark workloads and observed an IPC improvement of up to 4.98% on a Limit Order Book workload, compared with the baseline configuration.
Through this project, I learned how to modify a cycle-accurate processor simulator, track instruction dependencies, implement criticality-based scheduling, and evaluate microarchitectural changes using performance metrics. It also helped me understand the differences between implementing a research idea and reproducing the complete system described in a paper."
| What You Mentioned | Why It Was Done (The Motivation) | Problems Faced & How Solved (The Reality) | Core Concept |
|---|---|---|---|
| "Long-latency loads stall the ROB" | Prevent pipeline stalls when waiting hundreds of cycles for DRAM. | The Reorder Buffer has finite slots (e.g. 256). When the head instruction waits for DRAM, dispatch freezes; solved by calculating the load address sooner. | ROB Head Blocking & DRAM Latency |
| "Dependency tracking at dispatch" | Know which instruction produces each register value without heavy bookkeeping. | Maintained a last_writer[] table mapping destination registers to instruction IDs, storing up to 2 predecessor IDs per instruction. |
Producer-Consumer Tracking |
| "Adaptive latency threshold at retirement" | Distinguish true DRAM misses from fast L1/L2 cache hits. | Can only measure total latency at retirement; used an exponential moving average (EMA) threshold so it automatically adapts to different workloads. | Dynamic Latency Thresholding |
| "Backward BFS slice construction" | Identify all ancestor instructions that compute the memory address. | Ran a bounded BFS (depth capped at 32) over predecessor IDs to avoid search overhead, and stored delinquent PCs for proactive tagging. | Critical Slice Identification |
| "ROB head liveness guarantee" | Prevent deadlocks and instruction starvation when prioritizing critical code. | In out-of-order CPUs, the ROB head must retire in program order. Step 1 of execute always issues the ROB head first if ready before any critical instructions. | Scheduler Starvation Prevention |
| "Limit Order Book 4.98% IPC boost" | Validate performance on real-world latency-sensitive systems. | AVL tree operations involve short, repetitive pointer chasing; CRISP quickly learned the address chain and consistently prioritized it. | Workload Sensitivity & Pointer Chasing |
Answer: On physical CPUs, the pipeline stages, issue scheduler, and Reorder Buffer are hardwired into silicon—you cannot change instruction scheduling policies or add internal dependency tracking tables. ChampSim is a cycle-accurate C++ simulator that models the exact pipeline stages, caches, and memory controllers of an out-of-order CPU. This allowed us to modify the execution scheduler and measure cycle-accurate IPC gains.
Answer: A critical slice is the chain of instructions needed to compute the memory address of a slow load. For example, if a load needs address
A, andAis calculated by adding an offset to registerR1, andR1came from an earlier base-pointer load, then those earlier instructions form the slice. If the CPU executes them sooner, the slow load can be sent to memory sooner, reducing the time the processor sits idle.
Answer: In an out-of-order processor, instructions must retire in program order from the head of the Reorder Buffer. If we only prioritized critical instructions, the instruction at the head of the ROB might never get an execution slot, causing the entire CPU to freeze. We solved this with a strict liveness guarantee: Step 1 of our execution stage unconditionally issues the ROB head if it is ready. Critical instructions are only prioritized after the ROB head is taken care of.
Answer: Limit Order Book uses an AVL tree, which involves structured, repetitive pointer chasing along tree branches. The instruction chains that calculate node pointers are short, predictable, and repeat frequently, so CRISP quickly learns which instructions to prioritize. In contrast, SSSP is a graph algorithm with unpredictable, irregular memory accesses, so the dependency chains change constantly and cannot be reliably predicted.
Answer: Instead of hardcoding a static number like 100 cycles, the simulator tracks a running exponential moving average of memory latency across all loads. A load is flagged as slow only if its latency is significantly higher than this running average. This allows the system to automatically adjust between workloads with fast cache hits and workloads with frequent DRAM trips.