IEEE Solid-States Circuits Magazine - Fall 2020 - 64

and forth from compute to storage.
As the amount of data grows larger,
the data bus not only can become
a bottleneck in terms of speed but
can also consume too much energy
to transfer the data to the comput-
ing units.
One big data systems following
conventional computer architecture
is the disaggregated solution in large
scale, as shown in Figure 8(a) [27].
Sharing computing power and storage
can provide versatile solutions for dif-
ferent data problems. Computing is
executed by a group of powerful serv-
ers composed of CPUs, DRAM, and
network connectors. The data storage
side is made up of large numbers of
SSD drives connected through NVMe
interfaces via NVMe over fabric (NVMeoF). Compared with a converged sys-
tem, the disaggregated system has

the following advantages: 1) compute
and storage scaled independently,
2) compute and storage resources
shared, 3) utilization improved, 4) a
grow-as-you-go model used based on
workload demand, and 5) storage ser-
vices centralized. The drawback is that
it requires a high-speed network to
reduce latency.
There are three types of interfaces
for SSDs: serial advanced technology
attachment (SATA), attached small
computer system interface (SAS),
and peripheral component intercon-
nect express (PCIe)/NVMe. The SATA
and SAS protocols were originally
designed for HDD, and some trans-
lation is needed for the protocol to
be able to write to flash. Only NVMe,
a flash-specific protocol based on
PCIe, can fully take advantage of
faster media flash SSDs. The very

Disaggregated

Server Node

NVMe-oF EBOF

Server Node

25-GbE

Server Node

Ethernet

Ethernet
Switch

100-GbE
Ethernet

NVMe-oF EBOF

Server Node

NVMe-oF EBOF

Server Node
NIC

CPU

NVMe-oF EBOF

Switch

DRAM

NVMeoF SSD

(a)
Application
Computation Device Interface

Block Device Interface

Computation Management

Open Channel FTL

Computation Driver

NVMe Driver

wide adaptation in server space also
makes PCIe uniquely expandable
from a few to 100 PCIe lanes in a sys-
tem. PCIe switches can further help
the expansion of SSD lanes. NVMe is
a faster protocol that enables remote
drives to appear local to a host and
enables large pools of storage to be
shared in a data center.
Besides the increased interface
speed to more quickly move data
between storage and computing, many
people are thinking of different ways
to solve the big data problem-crunch
the data near the storage and transfer
only the intermediate results to the
host to combine the results together.
Many companies have already made
attempts to move the database func-
tions near storage, such as Scaleflux
with field-programmable gate array
solutions [28] and Alibaba's vertical
stack hardware/software approach
[29]. A few common functions need
to be accelerated, such as 1) deep
learning for big data analytics, 2)
erasure coding for redundant array
of independent disks data protec-
tion, 3) data compression to reduce
data footprint, 4) deduplication, and
(5) encryption for security. As shown
in Figure 8(b), Alibaba [29] built a
software stack to support the com-
pute in storage with computation
device interface and management
drivers. For the filtering functions,
Alibaba illustrated 1.56 times band-
width im--provement and 71% latency
reduction using in-storage computing
SSDs. With more companies adopting
in-storage computation strategies,
there has been progress within the
standardization committee to reach
agreement on common protocols and
application programming interfaces
for the industry [30].

PCIe
In-Storage Computing SSD
(b)
FIGURE 8: (a). The disaggregated system with big servers on one side and a large NVMe-oF
Ethernet-connected bunch of flash (EBOF) on the storage side. The data bus can be 25-GbE
Ethernet or 100-GbE Ethernet connections [26]. (b). The Alibaba compute in storage architecture. Alibaba owns the software/driver to add extra interfaces for computation device
interface and management [29]. NIC: network interface controller.

FA L L 2 0 2 0

IEEE SOLID-STATE CIRCUITS MAGAZINE

Conclusions
NAND flash is used extensively as a
replacement for HDD. NAND physi-
cal scaling continued to 128 layers
this year, and some manufactur-
ers have announced going up to
170-190 layers. NAND logical scal-
ing has made a lot of products fea-
sible with 3D NAND QLC. There is

IEEE Solid-States Circuits Magazine - Fall 2020

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