Lab-1 Rowhammer Mitigations

In this lab assignment, you will implement rowhammer mitigations. First you will implement two previous rowhammer mitigations, Graphene (MICRO'20), that uses algorithmic tracking of aggressor rows with few counters stored in SRAM, and CRA (CAL'14) which uses exact tracking of aggressors with in-DRAM counters tracking all row activations. Further, you will design HYDRA (ISCA'22), an improvement on CRA to reduce its overheads. Overall, You will learn how to model and simulate these hardware defenses, understand the performance trade-offs of different design decisions, and potentially create a more efficient solution.

Task-A - Graphene [4 points]

The first task is to implement the Misra-Gries tracker from Graphene. A basic framework of the Misra-Gries tracker is available in mgries.h/c. You need to implement functions to check the tracker on a memory access, reset the tracker at appropriate times, and perform the mitigative action (refresh neighboring rows), when the time is right. We have implemented a function memsys_rh_mitigate in memsys.c for you to use for invoking the mitigative action.

You will have to complete the following specific tasks:

1. Instantiate Misra-Gries tracker in memsys.c (mgries_t)
   • Think: How large should the tracker be? How many trackers do you need? What should be the threshold?
2. Implement the function to access and update the tracker; mgries_access.
3. Reset the tracker every 64ms using mgries_reset (assume a 4Ghz Clock).
4. Use the previous memory access information in dram_acc_info to appropriately invoke mgries_access.
5. Invoke the mitigation memsys_rh_mitigate when needed.

Grading Criteria:

We will use the value of the statistics MGRIES-*_NUM_INSTALL and MSYS_RH_TOT_MITIGATE to award points. +4 points if the values match the correct answer (+-5% error is acceptable), otherwise 0.

Task-B - CRA [3 points]

The second task is to implement CRA. A basic framework for the in-DRAM counters is available in cra_ctr.h/c. For this task, we will assume that out of the 16GB memory, the first 15GB stores the data and the last 1GB only stores the counters. For the simulation, you will store the counter values in a CraCtr structure within memsys.c and appropriately insert accesses to the DRAM to simulate the extra memory activity for accessing the in-DRAM counters. The first part of the task involves implementing the counters functionally, and the second part involves adding a counter-cache (structure Ctrcache provided in ctrcache.c/h) to reduce the number of DRAM accesses for accessing the counters.

The first sub-part has the following tasks:

1. Instantiate the CRA counters in memsys.c (cra_t). Note these counters will provide you with the correct value but you still need to access DRAM to do timing simulation.
2. Implement the functions to read and write to the in-DRAM counters in cra_ctr.c.
   • Think: How should the rows in memory map to the counters in DRAM?
3. Use the previous memory access information in dram_acc_info to appropriately invoke cra_ctr_read/write.
4. Invoke the mitigation memsys_rh_mitigate when needed.

The second sub-part has the following tasks:

1. Instantiate the ctr-cache in cra (ctr_cache).
2. Use ctrcache_access in indramctr_read/write to check if a counter is present in the cache and ctrcache_install to install if not present.
   • Think: How do you handle reads and writes to the cache? When will writes propagate to memory?

Grading Criteria:

We will use the value of the statistics CRA_NUM_READS, CRA_NUM_WRITES and MSYS_RH_TOT_MITIGATE to award points. +4 points if the values match the correct answer (+-5% error is acceptable), otherwise 0.

Task-C - HYDRA [3 points]

The third task is to implement HYDRA, to further reduce the overheads of CRA. You will focus on two sources of overheads: (a) number of reads and writes to the counters in DRAM, (b) number of times the mitigation is invoked.

Grading Criteria:

The points for this task will be awarded based on your results for the CRA_NUM_READS, CRA_NUM_WRITES and MSYS_RH_TOT_MITIGATE.

Files for Submission.

You will submit one tar with three folders inside the tar: src_lab1_a, src_lab1_b, src_lab1_c, each containing the entire source code (src folder) for the particular task. Please make sure that your code compiles and runs on either the Ubuntu or Red-Hat machines provided by ECE or CoC. We will compile, run and evaluate your code on similar machines. You should also submit a 2 page report (submitted as PDF) capturing the high-level details of your implementation and tables capturing the key results for parts (A)(B)(C) averaged across all the workloads.

Collaboration / Plagiarism Policy

• You are allowed to discuss the problem and potential solution directions with your peers.
• You are not allowed to discuss or share any code with your peers, or copy code from any source on the internet. All work you submit MUST be your own.
• You are also not allowed to post the problems or your solutions on the online.
• All submitted code will be checked for plagiarism against other submissions (and other sources from the internet) and all violations will receive 0 on the assignment.

DRAM-RH-SIM v1.0

A fast DRAM simulator for evaluating Rowhammer defenses.

Instructions to use the Simulator

• Compile: To compile, use cd src/; make all. This will download the traces from Dropbox (almost 800MB), and compile the source code to create sim executable.
• Run; To run the simulator with a workload's execution trace (e.g. ../TRACES/bwaves_06.mat.gz) running on 4 Cores of a CPU, use `./sim ../TRACES/bwaves_06.mat.gz ../TRACES/bwaves_06.mat.gz ../TRACES/bwaves_06.mat.gz ../TRACES/bwaves_06.mat.gz > ../RESULTS/BASELINE.4C.8MB/bwaves_06.res``.
  • You can run the simulator with any of the traces in the ../TRACES/ folder with the same trace or different traces running on the 4 CPU cores.
  • ../scripts/runall.sh provides the command for running one benchmark and also all benchmarks in parallel.
• Get Statistics: To get statistics, you can check the output of the simulator. For the bwaves_06 run above, you can see ../RESULTS/BASELINE.4C.8MB/bwaves_06.res for statistics like SYS_CYCLES : 2155238427.
  • To report a statistic (SYS_CYCLES) for all benchmarks and several results folders, use cd ../scripts; perl getdata.pl -w lab1 -amean -s SYS_CYCLES -d ../RESULTS/BASELINE.4C.8MB <ANOTHER RESULTS FOLDER>.
  • getdata.pl has several other useful flags including -gmean for geometric mean, -n for normalizing results of one folder with another, -mstat to multiply all stats with a constant value and more.

Pinned Repository

https://github.com/nayannair/mirage-plus-plus