figure 9: Average throughput as a function of ecc strength. the
system used 256mB of DRAm and 1GB of flash.
1. 2
1
Relative bandwidth
0.8
0.6
0.4
0.2
0
0
10
20
30
BCH strength
(a)SPECWeb99
40
accessed contents would be located in regions composed
of reliable low latency SLC. In general, we found that variable ECC strength gracefully extended Flash lifetime, and
that the overhead of ECC is minimized with configurable
density. Combining all of our techniques, we saw an average 20× lifetime improvement relative to a system using
only a single ECC. We believe our findings are applicable
not only to Flash but also to emerging memory technology
devices such as PCRAM. 1
New memory technologies are creating opportunities for
increased performance and efficiency in a data center. These
disruptive technologies are forcing architects to rethink the
current system memory and storage hierarchy in a server.
50
Together, with server virtualization, we believe these memory devices will help realize the objective of building greener
data centers.
11. 2
1
Relative bandwidth
0.8
0.6
0.4
0.2
0
0
10
20
30
BCH strength
(b) dbt2
40
figure 10: normalized expected lifetime for a given access rate and
the point of total flash failure.
Programmable Flash memory controller BCH1 error correcting controller
1
0.1
Normalized lifetime
0.01
0.001
0.0001
References
11. bedeschi, f. et al. a multi-level-cell
bipolar-selected phase-change
memory. in Proceedings of the
International Solid-State Circuits
Conference (feb. 2008), 428–625.
2. binkert, n., Dreslinski, r., hsu, l.,
lim, k., saidi, a., reinhardt, s. the
m5 simulator: modeling networked
systems. IEEE Micro 26, 4 (jul./aug.
2006), 52–60.
3. chang, l.-P. on efficient wear-leveling
for large-scale flash-memory storage
systems. in 22nd ACM Symposium on
Applied Computing (ACM SAC) (2007).
4. chang, l.-P., kuo, t.-W. real-time
50
garbage collection for flash-memory
storage system in embedded
systems. ACM Trans. Embedded
Computing Systems 3, 4 (2004).
5. cho, t. et al. a dual-mode nanD flash
memory: 1-gb multilevel and high-performance 512-mb single-level
modes. IEEE J. Solid State Circuits
36, 11 (nov. 2001).
6. flex-onenanD. http://www.samsung.
com/global/business/semiconductor/
products/fusionmemory/Products_
flexonenanD.html.
7. fusion-io’s solid state storage—a
new standard for enterprise-class
reliability. http://www.fusionio.com/
PDfs/Whitepaper_solidstatestorage2.
pdf.
8. hutsell, W, bowen, j., ekker, n. flash
solid-state Disk reliability. http://
www.texmemsys.com/files/f000252.
pdf.
9. intel x18-m/x25-m sata solid
state Drive. http://download.intel.
com/design/flash/nand/mainstream/
mainstream-sata-ssd-datasheet.pdf.
10. kgil, t., mudge, t. flashcache: a
nanD flash memory file cache
for low power web servers. in
International Conference on
Compilers, Architecture and Synthesis
for Embedded Systems (2006).
11. kgil, t., roberts, D., mudge, t.
improving nanD flash based
Disk caches. in Proceedings of
the International Symposium on
Computer Architecture (ISCA) (2008).
12. leventhal, a. flash storage today.
ACM Queue (aug. 2008).
13. meisner, D., gold, b.t., Wenisch, t.f.
Powernap: eliminating server idle
power. ASPLOS (mar. 2009).
14. metarams DDr3 metasDram
Doubles memory capacity and
increases frequency of future
intel systems. http://www.
metaram.com/pdf/press/metaram_
DDr3_08_ 19_08.pdf.
15. moshayedi, m., Wilkison, P. enterprise
ssds. ACM Queue (aug. 2008).
16. onfi: open nanD flash interface.
http://www.onfi.org/index.html.
17. scaramella, j. enabling technologies
for Power and cooling. http://h71028.
www7.hp.com/enterprise/downloads/
thermal_logic.pdf.
18. serial ata 2. 6 specification. http ://
www.sata-io.org.
19. Solaris ZFS Administration Guide.
2008.
20. university of massachusetts trace
repository. http://traces.cs.umass.
edu/ index.php/storage/storage.
alpha1
alpha2
alpha3
0.00001
unifor m
Web Search1
W eb Search2
Fin a n cial1
Fin a n cia l2
exp1
David Roberts ( daverobe@umich.edu),
advanced computer architecture lab,
Department of cse,
university of michigan.
Taeho Kgil ( taeho.kgil@intel.com),
intel corporation.
Trevor Mudge ( tnm@eecs.umich.edu),
advanced computer architecture lab,
Department of cse,
university of michigan.
idle,
13 all components save energy rather than just memory
and disk.
We also showed that a Flash memory controller with
reliability support greatly improves Flash lifetime. We
found that the best configuration of a Flash memory
controller is largely dependent upon the access patterns
resulting from the application. For example, we found
that the typical workload with Zipf access behavior was
best served by a Flash configured such that the heavily
© acm 0001-0782/09/0400 $5.00
106 communicAtionS of the Acm | aPril 2009 | Vol. 52 | no. 4
