27
February 2015
M
emories
Enabling MLC for industrial / embedded
applications
Knowing the WAF associated with an indi-
vidual usage model and drive pre-condition-
ing enables the calculation of the overall drive
endurance expressed in Terabytes Written
(TBW) and rate drives.
ated with an individual usage model and drive pre-‐conditioning enables the
drive enduranc expressed i Terabytes Written (TBW) and rate drives.
TBW = Capacity GB 1000 x P/E Cycles WAF
n which an individual usage model’s WAF would be in the range of 5 for
ock based mapping architecture, it can be affordable to replace an SLC Flash
cles by an MLC providing 3K write erase cycles and get the same endurance
terms
of
TBW.
e mapping granularity is tuned to an individual systems’ access pattern, the
igher its drives’ endurance.
hyMap
TM
can save money if found out that now
LC was required before.
n
ocks containing sectors or pages with obsolete data must be consolidated to
is process is called garbage collection (GC). An efficient GC algorithm selects
contain most obsolete pages. The process is either performed in the
times.
storage capacity compared to the net capacity of the logical data space is
visioning (OP). This is necessary to provide extra blocks so that GC can
valid data stored in several used blocks. Also, depending on the amount of
ore efficiently as the share of valid data compared to invalid data of used
ta has to be copied to new locations and more free space can be provided
can utilize available NVM resources to improve performance and reduce
g unused capacity for over-‐provisioning.
ith it feature in terms of performance are shown in table xx:
Block Based Mapping
Hyperstone hyMap
TM
Read (MB/s)
Write (MB/s)
Read (MB/s)
Write (MB/s)
95
60
90
60
5.7
0.09 (20 IOPS)
3.1
4.0 (1000
IOPS)
75
25
70
25
Considering a scenario in which an individ-
ual usage model WAF would be in the range
of 5 for hyMap™ and 200 for a block-based-
mapping architecture, it can be affordable to
replace an SLC flash with 100K write-erase
cycles by an MLC, providing 3K write erase
cycles and getting the same endurance in
terms of TBW. That means, the better the
mapping granula ity is tuned to an individu l
system access pattern, the lower the WAF and
the higher its drive endurance. hyMap™ can
save money if it is found that now MLC can
be used where SLC was required before.
Performance optimization
At some point physical blocks containing
sectors or pages with obsolete data must be
consolidated to free-up storage space. This
process is called garbage collection (GC). An
efficient GC algorithm selec s thos blocks
first that contain most obsolete pages. The
process is either performed in the back-
ground or during idle times. The availabil-
ity of excess storage capacity compared to
the net capacity of the logical data space is
referred to as over-provisioning (OP). This is
necessary to provide ext a blocks so that GC
can consolidate fragmented valid data stored
in several used blocks. Also, depending on
the amount of OP space, GC can work more
effici ntly as the share of valid data compared
to invalid data of used blocks decreases, less
data has to be copied to new locations and
more free space can be provided per oper-
ation. hyMap™ can utilize available NVM
resources to improve performance and
reduce WAF by dynamically using unused
capacity for over-provisioning.
The effects of hyMap™ in terms of perfor-
mance are shown in table 1.
Data reliability and device-life extending
features
hyMap™ uses a patented wear levelling (WL)
algorithm that can be configured to the
needs of different flash technologies as well
as application requirements. WL is used to
systematically utilize all flash blocks of the
system equally in terms of consuming their
individual write-erase-cycle endurance bud-
get. It supports dy amic, static, and global
wear levelling.
Dynamic WL requires no copy-overhead but
alone would be limited to blocks not contain-
ing data. Static WL includes also those blocks
containing data. Static data is relocated if
need d. This WL activi y is triggered at pre-
defined threshold levels. Also these routines are
executed in the background and interrupted in
case of higher priority host commands.
Power Fail Robustness without
external DRAM
As soon as a power-down is recognized, the
controller is reset and the flash is write-pro-
tected. A log of all recent flash transactions
is kept. Should the latest data be corrupt, the
c ntroll r will recover the l t st valid entry
before that last failed write. All mapping
information is reliably stored on the flash and
therefore exter al DRAM is not needed.
The new firmware concept is targeted to mak-
ing MLC flash as reliable as possible. As two
logical MLC flash p ges are physically cor-
related, it is possible to destroy data of an older
page by writing another new one within the
same block (paired pages). hyMap™ applies a
Reliable Write feature to cope with this occur-
rence to make MLC power-fail safe. More-
over, it uses Safe Flash Handling in unstable
power supply situations and especially when
exposed to a sudden power-down in which
the last programming of pages might not be
reliable although a flash might have reported
successful writing.
n
consolidate fragmented valid data stored in several used blocks. Also, depending on the amount of
OP space, GC can work more efficiently as the share of valid data compared to invalid data of used
blocks decreases, less data has to be copied to new locations and more free space can be provided
per operation.
hyMap
TM
can utilize available NVM resources to improve performance and reduce
WAF by dynamically using unused capacity for over-‐provisioning.
The effects of
hyMap™
with it feature in terms of performance are shown in table xx:
Performance comparison for an SD card in SD3.0 UHS-‐I Mode using S8 with block based mapping and with
hyMap
TM
Firmware
Data reliability and device-‐life extending features
hyMap
TM
uses a patented wear levelling (WL) algorithm that can be configured to the needs of
different Flash technologies as well as application requirem nts. WL is used to systematically utilize
Block Based Mapping
Hyperstone hyMap
TM
Read (MB/s)
Write (MB/s)
Read (MB/s)
Write (MB/s)
SLC
Sequentiell
95
60
90
60
4K Random
5.7
0.09 (20 IOPS)
3.1
4.0 (1000 IOPS)
MLC
Sequentiell
75
25
70
25
4K Random
4.7
0.03 (7 IOPS)
3.3
3.9 (975 IOPS)
SLC: 2x nm, 8K page, 100 MHz DDR I/F, 2 CE DDP
MLC: 1x nm, 16K page, 100 MHz DDR I/F, 2 CE DDP
Table 1. Performance comparison for an SD card in SD3.0 UHS-I mode using S8 with
block-based mapping and with hyMap™ firmware
