IEEE Consumer Electronics Magazine - July 2015 - 93

Enterprise SSD Touch Rate Versus Response Time
100,000

1 TB 64 GB 4 GB 256 MB 16 MB 1 MB 64 kB

16 TB

10,000
Hi IOPS 4 kB
Transaction

100

Semiactive

1
10%

Cold-Active

One Day = 86,400 s
One Hour = 3,600 s

µs

µs
10

s
m

s
1

s
0
10

1,
00

,0
0

0,
00

Inactive

0.1%

eSSD

Near Line

Touch/Y

1,000

Response Time

FIGURE 4. Touch/Y and response time for reading for an enterprise SSD during read.

LIMITS ON TOUCH RATE

flash used in consumer applications generally uses smaller-block-size content
movements as well.
Figure 5 shows the touch rate for
enterprise-grade PCIe attached flash,
which is faster than SATA or SAS SSDs.
Again, this curve is for read operations.
PCIe flash outperforms traditional SSDs
across the board, with an approximately
two-times faster response time and a
two-times greater touch rate at common
block sizes.
The touch rate charts highlight and
help quantify the achievable benefits
from various flash storage devices.

The touch rate charts shown for flash storage were for read operations only. The
write performance generally differs from
the read performance for flash storage
(i.e., it is often slower). The slower writes
are the result of the erase/write requirement for writing flash memory and may
be exaggerated by flash garbage collection and flash wear leveling.
One key restriction for flash write performance is the result of the limited write
endurance of the flash devices. This
restriction is commonly expressed for
flash devices as a drive writes per day

PCIe Flash Touch Rate Versus Response Time
16 TB 1 TB 64 GB 4 GB 256 MB 16 MB 1 MB

10,000

64 kB

1,000

Hi IOPS 4 kB
Transaction

100

Semiactive

1
10%

Cold-Active

One Day = 86,400 s
One Hour = 3,600 s

µs
10

µs

m
1

10
0

s
m

m
s

s
1

10
0

s
10
0

s
00

00
0

10
,

00
0

Inactive

1,
0

1%
0.1%

PCI-E

Near Line

Touch/Y

100,000

10
0,

the application being considered. Similarly, when comparing different systems for
a given application, the touch rate and
response time values should be compared
at the same object size.
So long as the system has enough
bandwidth to operate all of the devices
simultaneously, the touch rate curve is
independent of the number of storage
devices being used. Thus, the touch rate is
scale independent. As a consequence, the
touch rate curve is the same for one device
as it is for a system of such devices.
If the system has resource limits
(such as the available network bandwidth), then full scaling may not be
achieved and the touch rate could be
lower. If the data are striped in parallel
across the devices, the result is a shifting
of the object size points along the curve.
Striping increases the data rate (and,
thus, decreases the response time) but
decreases the effective block size on a
device (i.e., it spreads the block over
multiple devices). It can be thought of as
moving the object size locations on the
curve toward the right and down. Caching can also alter the system performance. The device charts also do not
include system caching effects.
There may be separate curves for
read and write operations. For example,
flash memory devices tend to have different performance characteristics for
reading and writing information.
The touch rate chart is useful for
determining other system behaviors. For
example, the response time for a 1-TB
object here is 1.7 h. This means that a full
drive scan on this 4-TB drive will take at
least 1.7 × 4 = 7 h to read or write the full
capacity of the drive. There are many
tasks that do this, including replication,
RAID rebuild, and full data scrub.
Figure 4 shows the touch rate for
enterprise-grade SSDs when reading.
The value of SSDs is readily apparent in
the response time shift for small-block
I/O. The 4-kB response time is 65 times
faster than the performance HDD and has
a 33-times greater touch rate. The saturation touch rate of 27,000/Y is only about
2.4-times better than the performance
HDD. Thus, an SSD is used primarily for
small-block-size workloads. Although
this analysis was for enterprise SSDs, the

Response Time

FIGURE 5. Touch/Y and response time for reading for an enterprise PCIe device.

july 2015

IEEE ConsumEr ElECtronICs magazInE

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - July 2015