Ramulator: A DRAM Simulator
Ramulator is a fast and cycle-accurate DRAM simulator [1, 2] that supports a wide array of commercial, as well as academic, DRAM standards:
- DDR3 (2007), DDR4 (2012)
- LPDDR3 (2012), LPDDR4 (2014)
- GDDR5 (2009)
- WIO (2011), WIO2 (2014)
- HBM (2013)
- SALP 
- TL-DRAM 
- RowClone 
- DSARP 
The initial release of Ramulator is described in the following paper:
Y. Kim, W. Yang, O. Mutlu. "Ramulator: A Fast and Extensible DRAM Simulator". In IEEE Computer Architecture Letters, March 2015.
For information on new features, along with an extensive memory characterization using Ramulator, please read:
S. Ghose, T. Li, N. Hajinazar, D. Senol Cali, O. Mutlu. "Demystifying Complex Workload–DRAM Interactions: An Experimental Study". In Proceedings of the ACM International Conference on Measurement and Modeling of Computer Systems (SIGMETRICS), June 2019 (slides). In Proceedings of the ACM on Measurement and Analysis of Computing Systems (POMACS), 2019.
 Kim et al. Ramulator: A Fast and Extensible DRAM Simulator. IEEE CAL
 Ghose et al. Demystifying Complex Workload–DRAM Interactions: An Experimental Study. SIGMETRICS 2019.
 Kim et al. A Case for Exploiting Subarray-Level Parallelism (SALP) in DRAM. ISCA 2012.
 Lee et al. Tiered-Latency DRAM: A Low Latency and Low Cost DRAM Architecture. HPCA 2013.
 Seshadri et al. RowClone: Fast and Energy-Efficient In-DRAM Bulk Data Copy and Initialization. MICRO 2013.
 Chang et al. Improving DRAM Performance by Parallelizing Refreshes with Accesses. HPCA 2014.
Ramulator supports three different usage modes.
- Memory Trace Driven: Ramulator directly reads memory traces from a file, and simulates only the DRAM subsystem. Each line in the trace file represents a memory request, with the hexadecimal address followed by 'R' or 'W' for read or write.
- 0x12345680 R
- 0x4cbd56c0 W
- CPU Trace Driven: Ramulator directly reads instruction traces from a file, and simulates a simplified model of a "core" that generates memory requests to the DRAM subsystem. Each line in the trace file represents a memory request, and can have one of the following two formats.
<num-cpuinst> <addr-read>: For a line with two tokens, the first token represents the number of CPU (i.e., non-memory) instructions before the memory request, and the second token is the decimal address of a read.
<num-cpuinst> <addr-read> <addr-writeback>: For a line with three tokens, the third token is the decimal address of the writeback request, which is the dirty cache-line eviction caused by the read request before it.
- gem5 Driven: Ramulator runs as part of a full-system simulator (gem5 ), from which it receives memory request as they are generated.
For some of the DRAM standards, Ramulator is also capable of reporting power consumption by relying on either VAMPIRE  or DRAMPower  as the backend.
 The gem5 Simulator System.
 Ghose et al. What Your DRAM Power Models Are Not Telling You: Lessons from a Detailed Experimental Study. SIGMETRICS 2018.
 Chandrasekar et al. DRAMPower: Open-Source DRAM Power & Energy Estimation Tool. IEEE CAL 2015.
Ramulator requires a C++11 compiler (e.g.,
Memory Trace Driven
$ cd ramulator $ make -j $ ./ramulator configs/DDR3-config.cfg --mode=dram dram.trace Simulation done. Statistics written to DDR3.stats # NOTE: dram.trace is a very short trace file provided only as an example. $ ./ramulator configs/DDR3-config.cfg --mode=dram --stats my_output.txt dram.trace Simulation done. Statistics written to my_output.txt # NOTE: optional --stats flag changes the statistics output filename
CPU Trace Driven
$ cd ramulator $ make -j $ ./ramulator configs/DDR3-config.cfg --mode=cpu cpu.trace Simulation done. Statistics written to DDR3.stats # NOTE: cpu.trace is a very short trace file provided only as an example. $ ./ramulator configs/DDR3-config.cfg --mode=cpu --stats my_output.txt cpu.trace Simulation done. Statistics written to my_output.txt # NOTE: optional --stats flag changes the statistics output filename
Requires SWIG 2.0.12+, gperftools (
libgoogle-perftools-devpackage on Ubuntu)
$ hg clone http://repo.gem5.org/gem5-stable $ cd gem5-stable $ hg update -c 10231 # Revert to stable version from 5/31/2014 (10231:0e86fac7254c) $ patch -Np1 --ignore-whitespace < /path/to/ramulator/gem5-0e86fac7254c-ramulator.patch $ cd ext/ramulator $ mkdir Ramulator $ cp -r /path/to/ramulator/src Ramulator # Compile gem5 # Run gem5 with `--mem-type=ramulator` and `--ramulator-config=configs/DDR3-config.cfg`
By default, gem5 uses the atomic CPU and uses atomic memory accesses, i.e. a detailed memory model like ramulator is not really used. To actually run gem5 in timing mode, a CPU type need to be specified by command line parameter
Ramulator will report a series of statistics for every run, which are written
to a file. We have provided a series of gem5-compatible statistics classes in
Memory Trace/CPU Trace Driven: When run in memory trace driven or CPU trace
driven mode, Ramulator will write these statistics to a file. By default, the
filename will be
DDR3.stats). You can write
the statistics file to a different filename by adding
--stats <filename> to
the command line after the
--mode switch (see examples above).
gem5 Driven: Ramulator automatically integrates its statistics into gem5.
Ramulator's statistics are written directly into the gem5 statistic file, with
ramulator. added to each stat's name.
NOTE: When creating your own stats objects, don't place them inside STL containers that are automatically resized (e.g, vector). Since these containers copy on resize, you will end up with duplicate statistics printed in the output file.
Reproducing Results from Paper (Kim et al. )
Debugging & Verification (Section 4.1)
For debugging and verification purposes, Ramulator can print the trace of every
DRAM command it issues along with their address and timing information. To do
so, please turn on the
print_cmd_trace variable in the configuration file.
Comparison Against Other Simulators (Section 4.2)
For comparing Ramulator against other DRAM simulators, we provide a script that
automates the process:
test_ddr3.py. Before you run this script, however, you
must specify the location of their executables and configuration files at
designated lines in the script's source code:
- DRAMSim2 (https://wiki.umd.edu/DRAMSim2):
- USIMM (http://www.cs.utah.edu/~rajeev/jwac12):
- DrSim (http://lph.ece.utexas.edu/public/Main/DrSim):
- NVMain (http://wiki.nvmain.org):
Please refer to their respective websites to download, build, and set-up the other simulators. The simulators must to be executed in saturation mode (always filling up the request queues when possible).
All five simulators were configured using the same parameters:
- DDR3-1600K (11-11-11), 1 Channel, 1 Rank, 2Gb x8 chips
- FR-FCFS Scheduling
- Open-Row Policy
- 32/32 Entry Read/Write Queues
- High/Low Watermarks for Write Queue: 28/16
test_ddr3.py <num-requests> to start off the simulation.
Please make sure that there are no other active processes during simulation to
yield accurate measurements of memory usage and CPU time.
Cross-Sectional Study of DRAM Standards (Section 4.3)
Please use the CPU traces (SPEC 2006) provided in the
cputraces folder to run
CPU trace driven simulations.
For estimating power consumption, Ramulator can record the trace of every DRAM
command it issues to a file in DRAMPower  format. To do so, please turn
record_cmd_trace variable in the configuration file. The resulting
DRAM command trace (e.g.,
cmd-trace-chan-N-rank-M.cmdtrace) should be fed
into a compatible DRAM energy simulator such as
VAMPIRE  or
DRAMPower  with the correct configuration
(standard/speed/organization) to estimate energy/power usage for a single rank
(a current limitation of both VAMPIRE and DRAMPower).
- Yoongu Kim (Carnegie Mellon University)
- Weikun Yang (Peking University)
- Kevin Chang (Carnegie Mellon University)
- Donghyuk Lee (Carnegie Mellon University)
- Vivek Seshadri (Carnegie Mellon University)
- Saugata Ghose (Carnegie Mellon University)
- Tianshi Li (Carnegie Mellon University)
We thank the SAFARI group members who have contributed to the initial development of Ramulator, including Kevin Chang, Saugata Ghose, Donghyuk Lee, Tianshi Li, and Vivek Seshadri. We also thank the anonymous reviewers for feedback. This work was supported by NSF, SRC, and gifts from our industrial partners, including Google, Intel, Microsoft, Nvidia, Samsung, Seagate and VMware.