Comparing Samsung’s Planar and V-NAND SSDs

Samsung 3D V-NAND Press PhotoSamsung recently introduced  its 3D V-NAND-based 850 SSD which, according to The Tech Report, uses the same MEX controller as the company’s 3-bit planar SSD, the 840, introduced last year.

Samsung said in its keynote speech at the 2013 Flash Memory Summit that V-NAND consumes an average of 27% less power and runs at least 20% faster than its planar counterpart in an SSD application, all while providing ten times the endurance.  It’s only natural to assume that this would allow designers to produce a V-NAND SSD that would significantly outperform its planar NAND counterpart.

The SSD Guy had an opportunity to review one of the new 850 SSDs, and solicited the help of Calypso Systems to give it a thorough test.

For those  who don’t know Calypso, this is the company that was behind the Storage Network Industry Association, or SNIA’s, solid state storage Performance Test Specification, or PTS.  This is the high end of SSD testing, and it is used by nearly all SSD makers as a means of comparing every aspect of their SSDs against those of their competition.

Calypso ran the full suite of tests on Samsung’s V-NAND 850 SSD and the TLC-based 840 SSD.  Both SSDs are 256GB models, and both use the same controller.  During the tests something interesting popped up which can be easily seen in the following two charts.

These charts plot the power each SSD consumes during a test of sequential 128KB read accesses followed by random IOPS.  Note that you can click on any of these charts for a larger view, and if you right-click you can open them in a different window, allowing you to use Alt-Tab to bounce back and forth between this text and the chart.

On these charts the read power is the red line measured on the right vertical axis, and write power is the green line, measured on the same vertical axis.  The block size for the test is shown as columns which are measured on the left vertical axis.  The X-axis gives elapsed time.  All tests are performed following the standard SNIA Performance Test Specification IOPS Test protocol.

Samsung SSD 850-Block Size & Power over Time - Calypso SystemsThe first chart shows the V-NAND 850’s power consumption.

For the 850 the sequential read power for a constant 128-byte block size is roughly 2.7 Watts.  When the test switches to random accesses of variable block sizes the SSD’s power consumption jumps to a relatively consistent 3.2 Watts for writes, while read power varies between 2.5 – 3.2 Watts, depending on the block size.

 

Samsung SSD 840-Block Size & Power over Time - Calypso SystemsThe second chart shows the same tests being performed on the 840.  This TLC-based SSD has a very different power profile.  Sequential read power for a constant 128-byte block size averages about 3.4 Watts.  When the test switches to variable block size random accesses the SSD’s power consumption drops to a steady 2.7 Watts for writes and a range of 1.6 – 2.6 Watts for reads.

It’s interesting that the behavior of these two drives differs in this way: For sequential writes, the 840 consumes more power than the 850 (the 840 consumes 3.4 Watts compared to the 850’s 2.7 Watts) while for random I/Os the 840 consumes less power than the 850 (the 840 uses up to 2.7 Watts compared to the 850’s consumption of as much as 3.2 Watts ).  In other words, the amount of power consumed for sequential vs random I/Os is flipped between the two drives: The 850 has higher sequential than random power, while the 840 has lower sequential than random power.  Overall, the 850 actually uses slightly more power than the 840, a fact that is more than offset by its higher performance.

Calypso believes that this may be due to a more aggressive garbage collection algorithm in the 850 than in the 840.  Samsung says MLC V-NAND has endurance similar to that of SLC planar NAND.  It would make sense to take advantage of this higher endurance by using a more aggressive garbage collection routine.

The performance measurements of the two SSDs also argue that the 850 may use more aggressive garbage collection algorithms than the 840.  The 850 provides significantly higher IOPS than the 840 in the test suite, as well as faster average and maximum response times.  The following three tables help to illustrate that fact.

The first table compares IOPS across all of the tests for the two SSDs for read/write mixes of 100% writes, 65:35% read/writes, and 100% reads.

 

 
Percent of Reads
 
0%
65%
100%
Block Size
840
850
840
850
840
850
512B
4,662
7,902
12,905
20,687
115,105
116,817
4KiB
4,920
8,390
12,800
21,569
76,551
74,353
8KiB
2,422
4,208
6,279
10,744
47,282
48,766
16KiB
1,232
2,108
3,102
5,351
27,848
26,906
32KiB
600
1,057
1,596
2,673
15,688
15,612
64KiB
305
527
769
1,322
8,201
8,288
128KiB
152
263
390
659
4,015
4,267
1,024KiB
19
33
86
89
475
543

Clearly the 850 performs consistently better than the 840 at all block sizes for constant streams of writes and for a 65:35 read/write mix.  At the 100% read workload the performance of the two SSDs is roughly equivalent.

The fact that the 850’s performance  advantage is stronger under higher write loads may stem in part from its use of V-NAND.  This technology uses a charge trap to store the bits instead of the floating gate used by standard planar NAND.  As explained in a post by The Memory Guy, a charge trap (which is required to make a 3D NAND) provides faster programming at lower power than does a floating gate.  This appears to be the basis for Samsung’s claims of faster speed and lower power.

IOPS isn’t the most important performance measurement for an SSD in most applications, though.  Response times are of perhaps a greater importance, as was shown in the Objective Analysis and Coughlin Associates report titled How Many IOPS Do You Really Need?

Our next table compares the average response times in milliseconds for the two SSDs across the same range of tests used to produce the charts and the table above.  Lower response times are better.

 

 
Percent of Reads
 
0%
65%
100%
Block Size
840
850
840
850
840
850
512B
6.86
4.05
2.48
1.55
0.28
0.27
4KiB
6.50
3.81
2.51
1.49
0.42
0.43
8KiB
13.21
7.60
5.12
2.98
0.68
0.66
16KiB
25.96
15.18
10.36
6.01
1.15
1.19
32KiB
53.36
30.26
20.13
12.05
2.04
2.05
64KiB
104.96
60.70
41.65
24.41
3.90
3.86
128KiB
208.68
121.54
82.09
49.04
7.97
7.50
1,024KiB
1,664.48
962.23
392.91
359.76
67.51
58.88

The 850 gives consistently faster response times across all block sizes for all writes and a 65:35 read/write balance when compared against the 840.  At 100% reads the two drives are roughly equivalent.

When maximum response times are compared in the table below, the 850 looks even better.  The maxim response times for the 850 (in milliseconds) are consistently much faster for 100% writes and a 65:35 read/write balance than those of the 840.  At smaller block sizes the 850 is two orders of magnitude faster in these columns.

 

 
Percent of Reads
 
0%
65%
100%
Block Size
840
850
840
850
840
850
512B
1,775.32
11.54
1,480.16
15.48
1.46
2.47
4KiB
1,384.64
12.54
1,172.20
16.12
2.16
2.23
8KiB
1,426.89
15.73
997.17
17.72
1.81
2.23
16KiB
1,307.78
29.28
1,299.88
23.99
2.26
4.10
32KiB
1,118.03
50.81
2,042.20
38.56
2.96
6.73
64KiB
1,711.53
82.98
1,376.26
77.29
4.68
7.40
128KiB
2,402.16
155.53
2,246.72
113.22
18.35
14.68
1,024KiB
6,998.16
4,067.86
3,594.95
1,686.92
365.41
255.80

At 100% reads the 840 slightly outperforms the 850 for all block sizes except for the largest two: 128 and 1,024KiB.

The 850’s superior response times would imply the use of better garbage collection algorithms, along with write speed improvements provided by V-NAND’s charge trap technology.

When you consider that the 850’s write performance is significantly superior to that of the 840, even though their power consumption is relatively similar, you might conclude that the 850’s design team decided to take advantage of the higher performance and endurance, and the lower power of the V-NAND to pack more SSD performance into the same power envelope as its planar counterpart.  The 850 does, in fact, use 18% less power than the 840 for a sequential 100% write load of 128KiB blocks, but the 850 uses 18% more power for random I/Os than the 840.  It appears that the 850 uses that modest power increase to dramatically increase write IOPS, and to lower average and maximum response times.

The SSD Guy should note that there are many contributors to SSD power consumption other than the NAND flash, including a power-hungry DRAM and the SSD controller, but in the end we have two SSDs based on the same controller that perform very differently most likely because of firmware adapted to harness the  power advantage provided by V-NAND.

The plots used in this post were automatically built in the Calypso CTS software using Calypso data analytics tools, and the data in the tables was excerpted from standard reports generated by the same tools. The Calypso Data Dashboard provides automated data management, visualization and analytics for test results analysis and comparison. All test data was taken on the Calypso RTP 4.0 SSD Reference Test Platform.

One Response to Comparing Samsung’s Planar and V-NAND SSDs

Leave a Reply

Your email address will not be published. Required fields are marked *