IEEE Solid-State Circuits Magazine - Spring 2016 - 73

CPU

Memory

MySQL

Aggregate

Scan
Heavy Traffic!

I/O Interface

Memory

Early Filtering

CPU
PIM

Scan
Intelligent SSD

Regular SSD

(a)

(b)

FIGURE 8: (a) Process-in-memory (e.g., data copy) and (b) in-storage-computing (e.g., ffiltering).

hyperscale operators have pioneered
new techniques based on commodity
hardware and open source software.
For example, in Hadoop, software
configurations are used to compensate for the low reliability of commodity hardware and performance
relies on the sequential bandwidth of
hard disk drives for batch analytics
processing. As the demands on realtime analytics have increased and
the cost of solid-state drives (SSDs)
has fallen, SSD adoption in the data
center has grown substantially.
SSDs were originally introduced
as a drop-in replacement for hard

disk drives, keeping the legacy
software stack and hardware interfaces in place. However, substantial
progress has been made in updating systems and software to better support flash and complement
degrading NAND flash parameters.
For example, multistream, listed
as a T10 standard, can reduce
garbage collection overhead and
write amplification significantly [9].
(See Figure 9.)
Fusion IO was the first company
to use a PCIe interface on an SSD,
followed by the 2011 introduction of
the SDD-optimized, NVMe protocol.

1.2
Multistreamed SSD
1
0.8
0.6
Traditional SSD

0.4
0.2

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

Update Throughput (ops/s)

connectivity, and form factor, which
has driven demand for more customized solutions and limited the economies of scale for hardware manufacturing. IoT platforms combine components
in various configurations, but we expect that eventually edge hardware
will mostly converge to a single die or
package where a microcontroller unit
(MCU), memory, and storage are embedded together.
Typically, generated data are
collected at a local gateway before
being sent to the data center over
the network. Because raw data tends
to be large and noisy, the gateway
will typically filter and refine the
data before transmission to save
both networking bandwidth and data
center resources. Doing some of this
prefiltering and early data management thorough processing-in-memory and in-storage-computing has
been shown to reduce total power
consumption and processing time.
(See Figure 8.) While these technologies are still early in their development, we expect that, as they mature,
their adoption and demand will be
enhanced through management of
IoT data. Finally, at the data center,
collected data are stored and processed to extract value.
While traditionally data centers
have been used for business intelligence, deploying high-performance
hardware with proprietary software,

Time (min)

FIGURE 9: Update throughput comparison: Cassandra database with Yahoo cloud serving
benchmark [9].

IEEE SOLID-STATE CIRCUITS MAGAZINE

S P R I N G 2 0 16

Table of Contents for the Digital Edition of IEEE Solid-State Circuits Magazine - Spring 2016