A Case for Small Row Buffers in Non-Volatile Main Memories
Justin Meza
∗
Jing Li
†
Onur Mutlu
∗
∗
Carnegie Mellon University
†
IBM T.J. Watson Research Center
{meza,onur}@cmu.edu jli@us.ibm.com
Abstract—DRAM-based main memories have read operations that
destroy the read data, and as a result, must buffer large amounts of data
on each array access to keep chip costs low. Unfortunately, system-level
trends such as increased memory contention in multi-core architectures
and data mapping schemes that improve memory parallelism lead to only
a small amount of the buffered data to be accessed. This makes buffering
large amounts of data on every memory array access energy-inefficient;
yet organizing DRAM chips to buffer small amounts of data is costly, as
others have shown [11].
Emerging non-volatile memories (NVMs) such as PCM, STT-RAM,
and RRAM, however, do not have destructive read operations, opening
up opportunities for employing small row buffers without incurring
additional area penalty and/or design complexity. In this work, we
discuss and evaluate architectural changes to enable small row buffers
at a low cost in NVMs. We find that on a multi-core system, reducing
the row buffer size can greatly reduce main memory dynamic energy
compared to a DRAM baseline with large row sizes, without greatly
affecting endurance, and for some NVM technologies, leads to improved
performance.
I. INTRODUCTION
Modern main memory is composed of dynamic random-access
memory (DRAM). A DRAM cell stores data as charge on a capacitor.
Over time, this charge leaks, causing the stored data to be lost. To
prevent this, data stored in DRAM must be periodically read out and
rewritten, a process called refreshing. In addition, reading data stored
in a DRAM cell destroys its state, requiring data to be later restored,
leading to increased cell access time and energy. For this reason,
DRAM devices require buffering data which are read. To keep costs
low, the buffering circuitry in DRAM devices is amortized among
large rows of cells, in peripheral storage called the row buffer, at
least one per bank [2]. Refreshing data and buffering large amounts of
data wastes energy in DRAM devices, causing main memory power
to constitute a significant fraction of the total system power.
Data fetched into the row buffer, however, can be accessed at much
lower latencies and less energy than accessing the DRAM memory
array. Therefore, large row buffer sizes can improve performance
and efficiency if many accesses can be served in the same row.
Unfortunately, there are several reasons why such row buffer locality
can be low in systems: (1) some applications inherently do not have
significant locality within rows (e.g., random access applications),
(2) as more cores are placed on chip, applications running on those
cores interfere with each other in the row buffers, leading to reduced
locality, especially if the memory scheduling policy is unaware of
applications’ interference in the row buffers [7], as also observed
by others [10, 11], and (3) interleaving techniques that improve
parallelism in the memory system (e.g., cache block interleaving)
tend to reduce row buffer locality because they stripe consecutive
cache blocks across different banks. As a result, the performance
benefit of large row buffers may decrease in multi-core systems.
New non-volatile memory (NVM) technologies, such as phase-
change memory (PCM), spin-transfer torque RAM (STT-RAM), and
resistive RAM (RRAM), on the other hand, provide non-destructive
reads and do not require refreshing and restoring their data after
sensing. This is because NVMs do not store their data as charge, and
thus their data persists after being read. This not only eliminates the
refresh problem of DRAM devices but also opens up opportunities for
employing smaller row buffers in NVMs without incurring additional
area penalty and/or design complexity.
II. MOTIVATION
Emerging NVM technologies have several promising attributes
compared to existing memory technologies such as SRAM (used in
on-chip caches), DRAM, and Flash. For example, NVMs provide cost
advantages compared to SRAM and DRAM, and latency advantages
compared to Flash. Importantly, these NVMs feature non-destructive
read operations, which DRAM does not have (i.e., data sensing does
not destroy the contents of cells).
Typical DRAM chip micro-architectures (JEDEC-standard DDR-
type SDRAM) are divided into banks that consist of rows (wordlines)
and columns (bitlines). Due to physical pin limitations, all the
information required to service a memory request must be supplied
over multiple commands. The Row Address Strobe (RAS) command
sends the row and bank address to select one of the banks and a row
within that bank. Then, an entire row (usually 1 to 2KB per chip)
is read out into latch-based sense amplifiers which comprise the row
buffer [2]. The Column Address Strobe (CAS) command then selects
a subset (i.e., column) of data from the row buffer (8B in a DDR3
×8 device [6]). Thus, a DRAM access first fetches many kilobytes
of data into the row buffer (RAS) and, in the worst case, uses only
a tiny portion of it (CAS). If multiple columns of the row buffer are
needed, multiple consecutive CAS commands can be issued, which
amortizes the cost of fetching the large row into the row buffer.
To illustrate how much buffered data is actually used in real
applications, Figure 1a shows average row buffer locality (row hit
rate) when employing various row buffer sizes on several system con-
figurations using the FR-FCFS scheduling policy [8].
1
In particular,
we show 1- and 8-core systems employing two different schemes
for mapping data in main memory: (1) row interleaving, which
places consecutive memory addresses in the same row, and (2) block
interleaving, which stripes data in consecutive memory addresses
(usually cache blocks) across different banks. Row interleaving helps
exploit row buffer locality by enabling data with spatial locality
to reside in the same row buffer, while block interleaving aims
to improve memory parallelism by enabling concurrent access of
memory channels/banks for consecutive memory addresses.
Comparing the 1- and 8-core row-interleaved data, we see that
while row interleaving does enable more row buffer locality, its
benefits diminish as memory system contention increases with more
cores: row buffer hit rate is less than 50% for row interleaving even
with large, 1KB rows. Block interleaving reduces row buffer locality
over row interleaving, to less than 10% in the 8-core case. While
it is clear that row locality is lower on multi-core systems, what
is less obvious is how row buffer size affects system-level tradeoffs,
such as energy-efficiency, performance, and durability, in NVM main
memories. This work evaluates these tradeoffs.
8B 16B 32B 64B 128B 256B 512B
1KB
Per DRAM-Chip Row Buffer Size
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Row Buffer Locality
1-core, row interleaved
8-core, row interleaved
1-core, cache block interleaved
8-core, cache block interleaved
(a)
Bank N
Bank 2
Bank 1
Bank 0
Memory Array
Row Buffer
Column Decoder
and Multiplexer
Row Decoder
...
...
...
(b)
Fig. 1: Row size affects row locality (a); our NVM architecture (b).
1
Application-aware memory request scheduling policies (e.g., [1, 3, 7])
provide better performance, but they can reduce row buffer locality.
1
8B 16B 32B 64B 128B 256B 512B
1KB
Baseline
Per DRAM-Chip Row Buffer Size
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
Normalized Memory Energy
STT-RAM
DRAM
PCM
(a) Memory energy with block interleaving.
8B 16B 32B 64B 128B 256B 512B
1KB
Per DRAM-Chip Row Buffer Size
0
1
2
3
4
5
Weighted Speedup
STT-RAM
DRAM
PCM
(b) Performance with block interleaving.
8B 16B 32B 64B 128B 256B 512B
1KB
Per DRAM-Chip Row Buffer Size
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
1.1
Normalized Writes
No cache
32MB cache
(c) Writes with and without a 32MB cache.
Fig. 2: Multi-core results for energy (normalized to DRAM with 1KB rows), performance, and number of writes (normalized to 1KB rows).
III. A SMALL ROW BUFFER NVM ARCHITECTURE
Figure 1(b) shows the organization of our NVM architecture. Com-
pared to a traditional DRAM organization, the physical placement of
the row buffer and the column multiplexer (part of the I/O gating
circuitry in DRAM designs) are swapped in the data path (shown in
gray). This rearrangement makes better use of resources by sharing a
smaller number of sense amplifiers (the devices which store bits in the
row buffer) among multiple bitlines. Note that this is not possible in
DRAM (without reducing the row size) because a sense amplifier for
each bit in the row is required in DRAM to restore the charge of the
cell after it is read. Unlike DRAM, however, our organization requires
decoding both the row address and the column address during a RAS
command, so that only a subset of the row containing the bits of
interest will be selected, sensed, and stored in the row buffer. During
a CAS command, the data bits from the row buffer corresponding to
the desired column are further selected by the I/O gating circuitry
and sent to a prefetch buffer.
2
While related prior work [4] employed multiple, narrow rows
in a PCM main memory for reducing array reads and writes, it
focused on (1) a traditional DRAM data path design, (2) an iso-
area reorganization, requiring more area overhead than our technique
which employs smaller row buffers, and (3) assumed a standard
DRAM protocol for device access.
IV. RESULTS
We developed a cycle-accurate DDR3 memory simulator which we
use as part of an in-house x86 multi-core simulator, whose front-end
is based on Pin. We modify our memory simulator timings according
to those in Table I for PCM and STT-RAM. We show results for an
8-core system with different memories and row buffer sizes, though
reducing row buffer size in DRAM incurs significant area overhead
and chip cost, as discussed in [10, 11], which we do not evaluate.
We evaluate 31 multiprogrammed workloads composed of SPEC,
TPC, and STREAM benchmarks. We will focus on a DRAM chip
micro-architecture with 1KB row buffers and block interleaving as
our baseline (our findings are similar for row interleaving [5]).
Technology Energy (Read/Write) Latency (Read/Write)
PCM 2×/100× 5×/10×
STT-RAM 0.5×/1× 1×/1×
TABLE I: NVM array parameters, relative to DRAM.
Energy (Figure 2a): In all cases, reducing the row buffer size can
significantly reduce memory energy consumption, though there are
diminishing marginal returns. The diminishing marginal returns are
because, as the row buffer size decreases, memory energy becomes
dominated by the energy required to transfer data between the row
buffer and I/O pads during read and write operations.
A modest row buffer size of 64B per chip leads to 47%/67%
less main memory energy for PCM/STT-RAM, compared to an all-
DRAM main memory with large rows (1KB per chip). Note that this
reduction is achieved despite worse underlying technology parameters
2
For more details, please refer to our accompanying tech report [5].
than DRAM (cf. Table I) because the energy saved by reducing the
row buffer size more than makes up for the higher average memory
array access energy. Hence, an NVM main memory with smaller
row buffers can significantly reduce memory energy consumption
compared to a DRAM baseline with large row buffers.
Performance (Figure 2b): We evaluate the performance of our
system using the weighted speedup metric [9] (higher is better). For
a given memory technology, reducing the row buffer size does not
greatly affect system performance due to the already low row buffer
locality present on our multi-core system (cf. Figure 1a). Interestingly,
with similar technology-dependent timing parameters as DRAM, an
STT-RAM main memory can achieve better performance because our
new organization enables a more efficient access protocol (detailed
in [5]) which eliminates the precharge delay incurred on row buffer
misses, and relaxes the t
RRD
and t
FAW
timing parameters to enable
more banks to be accessed simultaneously.
Durability (Figure 2c): NVM cells have a limited lifetime in terms
of the number of times they can be written to before their ability
to store data fails. We examine the effects of different row buffer
sizes on device durability with and without a small 32MB e-DRAM
cache to a PCM main memory. We find that with or without a cache,
decreasing the row buffer size has only a small effect on the number
of NVM writes performed due to the low row buffer locality present
in the system. In contrast, the addition of a reasonably-sized e-DRAM
cache has a large impact on the reduction of writes, decreasing the
number of writes by 39% to 47% across the various row buffer sizes.
V. CONCLUSIONS
We showed that on a multi-core system, reducing the row buffer
size can greatly reduce main memory dynamic energy compared
to a DRAM baseline with large rows, without greatly affecting
performance and durability. Our future work includes exploring
architectural techniques which effectively leverage small row buffer
sizes for improved performance and energy-efficiency.
REFERENCES
[1] Y. Kim et al. ATLAS: A scalable and high-performance scheduling
algorithm for multiple memory controllers. HPCA ’10.
[2] Y. Kim et al. A case for exploiting subarray-level parallelism (SALP)
in DRAM. ISCA ’12.
[3] Y. Kim et al. Thread cluster memory scheduling: Exploiting differences
in memory access behavior. MICRO ’10.
[4] B. C. Lee, E. Ipek, O. Mutlu, and D. Burger. Architecting phase change
memory as a scalable DRAM alternative. ISCA ’09.
[5] J. Meza, J. Li, and O. Mutlu. A case for small row buffers in non-volatile
main memories. http://safari.ece.cmu.edu/tr/tr-2012-002.pdf.
[6] Micron. 1Gb: ×4, ×8, ×16 DDR3 SDRAM data sheet. http://download.
micron.com/pdf/datasheets/dram/ddr/1GbDDRx4x8x16.pdf.
[7] S. P. Muralidhara et al. Reducing memory interference in multicore
systems via application-aware memory channel partitioning. MICRO’11.
[8] S. Rixner, W. J. Dally, U. J. Kapasi, P. Mattson, and J. D. Owens.
Memory access scheduling. ISCA ’00.
[9] A. Snavely et al. Symbiotic jobscheduling for a simultaneous multi-
threading processor. ASPLOS ’10.
[10] K. Sudan, N. Chatterjee, D. Nellans, et al. Micro-pages: increasing dram
efficiency with locality-aware data placement. ASPLOS ’10.
[11] A. N. Udipi, N. Muralimanohar, N. Chatterjee, et al. Rethinking DRAM
design and organization for energy-constrained multi-cores. ISCA ’10.
2