IEEE Solid-States Circuits Magazine - Fall 2020 - 62

read latency, and it is rarely used. In
system applications, scanning algo-
rithms can be used to monitor the
data retention situation. If the data
are found to have lost some margin,
they can be moved to another physi-
cal location before data loss occurs by
reading the data with optimum read.
Another important specification is
the PE cycle, which can also translate
to DWPD in SSDs. The paper analogy
is that paper may rip after writing
and erasing on the same spot many
times. PE cycling forces the electrons
or holes to tunnel through the tun-
neling oxide many times. The tunnel-
ing oxide can be damaged, with traps
created within. This is normally
reflected in the widened distribution
with larger noise factors. Fresh uncy-
cled blocks normally have larger Vt
margins than cycled blocks. To uni-
formly cycle NAND blocks, a wearleveling algorithm can be used. Each
block can have its erase cycle num-
ber recorded. When one block experi-
ences a large number of PE cycles, it
can be retired to avoid data loss.

Open Channel Versus
Zone Namespace
As mentioned in the previous sec-
tion, flash applications require active
management, where a dedicated con-
troller is used for flash management.
Because NAND is a block erase device,
the flash has to be managed by the
flash translation layer (FTL), where the
physical addresses are linked to logi-
cal addresses, as the physical address
may be changed during data move-
ments (defragmentation or garbage
collection and wear leveling [20]). As
shown in Figure 6(a), traditional flash
controllers manage block metadata,
data buffering, wear leveling, and
error handling tasks locally in the
SSD controller. All the housekeep-
ing tasks, such as garbage collection,
wear leveling, and background scan,
consume extra PE cycles and reduce
system performance. The extra PE
cycles consumed will be indicated
by the write amplification (WA),
which can be much greater than one,
the number for writing host data
only. The WA is defined by the fol-

Host System

Block Meta Data

SSD

lowing [21]: bytes written to NAND
(which include housekeeping tasks
plus host data) divided by the host
data. For example, WA = 4 means
that one host data write to NAND
will have three times the number of
flash writes for housekeeping tasks.
WA is normally measured when the
SSD drive is near full capacity to get
the worst-case senario, where the
data defragmentation involves many
block data consolidations. To avoid
accumulating too many housekeep-
ing tasks, a complicated schedul-
ing algorithm is used to spread the
housekeeping activity in the back-
ground uniformly.
Customers who have mission-crit-
ical tasks prefer controlling NAND
background activity by the host appli-
cation to avoid performance penalties
in peak hours, then scheduling house-
keeping in the off-peak hours. Only
host applications will know when
the peak performance is needed and
when the background operations can
be executed in off-peak hours. Open
channel is an idea to move most of

Host (ZNS)
Physical Addressing R/P/E
SSD

Block Metadata

Data Buffering

Block Metadata

Data Buffering

Wear Leveling

Data Buffering

Physical Addressing R/P/E
Wear Leveling

Wear Leveling
Open Channel SSD
Error Handling

Error Handling

Flash Interface

NAND Media

Traditional

Open Channel

ZNS

(a)

(b)

(c)

FIGURE 6: (a) A traditional flash controller that manages block metadata, data buffering, wear leveling, and error-handling tasks. (b) An open-channel SSD controller where only error handling and flash interfaces remain in the simple controller. The management of metadata, data buffering, and
wear leveling are done in the host. (c) The ZNS controller where the physical zones are open to the host. The controller does all of the conventional
data management in each zone. ZNS: zone name space.

FA L L 2 0 2 0

IEEE SOLID-STATE CIRCUITS MAGAZINE

IEEE Solid-States Circuits Magazine - Fall 2020

Table of Contents for the Digital Edition of IEEE Solid-States Circuits Magazine - Fall 2020