Understanding MFT
  
Free Space Important
  
Optimizing Drive Life
  
Assuring Drive Life
  
Journals and Logs
  
How fast is fast enough?
  
When is MFT less than perfect?
  
MFT for Laptops
  
MFT for Servers
  
MFT Pricelists
  
Drive Appraisals
  
Downloads
  
Literature
  
Technical Notes
  
Resellers

Flash Drives With MFT Are
the Cheapest Way to Achieve
High Storage Device Speed

The two simple measures of the cost and value of storage are the cost per gigabyte and the cost per random read or write IO per second (IOPS) from that device. Using both these measures, the following can be concluded:
1. MFT, with Multi-Level Cell (MLC) drives, is today less expensive per effective gigabyte than 15,000 rpm SAS drives, and operates at a speed typically 20 times faster, with a cost per IOPS that is in some cases less than 3% of the SAS drives.
2. MFT, with Multi-Level Cell (MLC) drives, costs only half per gigabyte of what high performance Single-Level Cell (SLC) Flash drives such as Mtron and MemoRight Cost.
3. MFT, with Multi-Level Cell (MLC) drives, perform typically seven to ten times faster than Mtron and MemoRight SLC drives perform, and has a cost per IOPS that is as much as 40 times lower than the IOPS costs of these products.
4.

MFT, when used with SLC drives, performs at about half the reported speed of a Zeus IOPS Enterprise grade intelligent flash drive. But the IOPS cost of the Zeus is reported to be similar to that for 15k SAS drive, and is thus ten times higher than the MFT based product prices.

5. While MFT with MLC is clearly usable for the 95% of servers that are general purpose, it is also clear that it is suitable for many applications regarded as Enterprise class, at a cost less than 10% of what Enterprise technology now goes for.
6. Given that MLC is forecast to dramatically decline in cost over the next three years, it is reasonable to expect that most of these statements will be even more-true at that time.

The following discussion and tables will explore these conclusions.

Baseline Hard Disk Costs

As a baseline, let's consider the issues of storage costs and IOPS cost from the perspective of hard drives first. Consider a 7200 rpm 2.5" commodity Laptop drive costing $150, having 200gb of storage, and having a random IOPS rate of about 70. Next, consider a 15,000 rpm SAS drive having a retail price of about $350, a storage capacity of 36gb, and a random IOPS rate of 250. As a single drive, the 7200 rpm drive can be said to have a cost per gigabyte of about $0.75 (i.e. $150/200gb). Similarly, it can be said to have an IOPS cost of $2.14 (i.e. $150/70 IOPS). By parallel reasoning, the 15,000 rpm SAS drive can be said to have a cost per gigabyte of $9.72 (i.e. $350/36gb) and a cost per IOPS of $1.40 (i.e. $350/250 IOPS).

The above analysis is valid for single-drive environments such as Laptops and Workstations. It is understated for multi-drive hard drive environments such as mirrored and striped RAID environments which cost more both per effective gigabyte of space and thus more per IOPS as well. For instance, as a rule of thumb, a Raid-10 environment with eight drives can reasonably be argued to have a cost per usable gigabyte of storage that is twice the price of the above, and a cost per IOPS that is 2.3 times as high. Similarly, a less expensive RAID-5 array of five drives can be argued to have a storage cost 1.25 times the above, and a cost per IOPS that is about 1.83 times as high. We mention this distinction because with flash drive technology, the same rules do not apply. RAID-5 is almost always the least expensive way to achieve maximum performance when using Flash media, as demonstrated in this article. (But, we digress).

Core Statistics for Selected Flash Drives

The following table considers the core statistics of a number of Flash drive models. Among these, there are three categories.

The first are the market prices and statistics for various Mtron and MemoRight models as these offer the greatest value among SLC drives. Other models such as the Samsung, OCZ, and SuperTalent lines, but as a general rule, while these will often have a lower cost per gigabyte, they will tend to have significantly higher costs per IOPS due both to somewhat slower random read speed and often significantly slower random write speed.

The second are various MLC drives. Here, SuperTalent and Mtron are typical of the performance level ranges we can expect, as well as the size formats likely to be present.

Finally, we list several other MLC alternatives. One of these is flash disks constructed from inserting CF cards into a SATA case. Another is constructed by inserting CF cards into a PCI card, a very easy assembly to install. The third is an extremely low-cost private label brand of MLC drives. This was chosen in part because it demonstrates how significantly poor random access speeds can impact the performance of a product. A number of Flash SSDs have poor access times, what ever their other strengths or weaknesses.

Basic Statistics of Various Readily Available Flash SSD Models
Model Capacity (GB) Drive
Market
Price
Linear Read mb/sec Linear Write mb/sec Random Access Time millisec
2.5" 15,000 rpm SAS Hard Disk 36 350 110 110 4.00
2.5" 7,200 rpm SATA Hard Disk 200 150 70 70 15.00
Mtron Mobi 32gb 32 643 110 80 0.04
Mtron Mobi 64gb 3.5" 64 1,216 110 80 0.04
Mtron Pro 32gb 32 805 120 90 0.04
Mtron Pro 64gb 64 1,517 120 90 0.04
MemoRight 32gb 32 879 100 100 0.10
MemoRight gt 64gb 64 1,989 120 120 0.10
MemoRight gt 128gb 128 3,539 120 120 0.10
SuperTalent 32gb MLC 30 299 120 40 0.50
SuperTalent 64gb MLC 60 479 120 40 0.50
SuperTalent 128gb MLC 120 669 120 40 0.50
Mtron Mobi-1000 32gb MLC 32 290 100 40 0.04
Mtron Mobi-1000 64gb MLC 64 476 100 40 0.04
MLC CF 32gb kit 32 179 35 17 0.20
Proprietary 32gb MLC 32 179 100 30 0.40
Addonics PCI MLC CF 64gb kit 64 329 35 17 0.20

Computing Random Read and Write IOPS

In the following table, we compute random read and write rates per second based upon the data in the preceding table, as well as data generated from our own testing and data ferreted out over the internet.

Data regarding random read times is easily computed. All one needs to do to compute the random reads of a particular size is to divide the linear speed by the size of the block, add in the access time, and then convert the total from a microseconds elapsed time to a quantity per second.

Similarly, the random write times for MFT are easy to compute. All one has to do is to divide the linear write speed by the block size. Test after test has established that MFT performs at the published linear write speed of the drive provided this is reported correctly, because MFT writes in erase block increments.

Regarding actual native performance random writes, results for Mtrons and for CF cards are based upon results determined by EasyCo. Results for Memo Right are based upon tests reported on the Internet. Results for SuperTalent are based upon known general limits of MLC Flash media, as well as inferences. Here, we may have been slightly too generous.

Computed or Reported Raw Performance of Various Flash Models
Model 8k Random Rds/sec 16k Random Rds/sec 8k Random Wts/sec 16k Random Wts/sec 8k Random Wts/sec w/ MFT 16k Random Wts/sec w/ MFT
15,000 rpm SAS Hard Disk 250 250 250 250 n/a n/a
7,200 rpm SATA Hard Disk 70 70 70 70 n/a n/a
Mtron Mobi 32gb 8,871 5,392 123 123 10,000 5,000
Mtron Mobi 64gb 3.5" 8,871 5,392 123 123 10,000 5,000
Mtron Pro 32gb 9,375 5,769 123 123 11,250 5,625
Mtron Pro 64gb 9,375 5,769 123 123 11,250 5,625
MemoRight 32gb 5,556 3,846 271 234 12,500 6,250
MemoRight gt 64gb 6,000 4,286 271 234 15,000 7,500
MemoRight gt 128gb 6,000 4,286 271 234 15,000 7,500
SuperTalent 32gb MLC 1,881 1,752 64 62 750 375
SuperTalent 64gb MLC 1,881 1,752 64 62 750 375
SuperTalent 128gb MLC 1,881 1,752 64 62 750 375
Mtron Mobi-1000 32gb MLC 8,333 5,000 25 25 5,000 2,500
Mtron Mobi-1000 64gb MLC 8,333 5,000 25 25 5,000 2,500
MLC CF 32gb kit 3,043 1,795 3 3 2,125 1,063
Proprietary 32gb MLC 2,083 1,786 15 15 3,750 1,875
Addonics PCI MLC CF 64gb kit 3,043 1,795 3 3 2,125 1,063

Use of Microsoft Exchange Server Benchmarks
as the Basis for Assessing Overall Performance

Microsoft publishes a detailed testing protocol for its Exchange email and collaboration Server. In the broadest sense, this test based upon common usage patterns, recognizes that about 30% of Exchange's activity is write activity, while 70% is read activity. Similarly, it recognizes an average read or write size of about 11,000 bytes, which can be expressed (more or less) as the average of two 8k elements of activity and one 16k element (average element size 10.66kb). All of our performance estimates have been based upon that standard and mix of reads/writes and data sizes.

While one might choose another standard for comparison, the key advantages of the Exchange standard are (a) that it applies to a very large number of servers in its own right, and (b) because many other servers which we might wish to consider (e.g. database servers) would result in close but less conservative numbers. Conservative numbers are always desirable to prevent under-engineering.

Consideration of the Performance Results and Costs of Bare Flash Drives

The following table shows the essential description and price information of the Flash Drives previously discussed. It then computes the average IOs per second, based upon the Windows Exchange Server measure.

From this, we see that the best of SLC flash drives - the Mtron and MemoRight SLCs - perform at IOP rates that are superior to the fastest Hard Disks by a significant measure. In the case of the Mtrons, the product is overall 60% faster, while in the case of the MemoRights, it is about three times as fast as a 15k rpm SAS drive.

On the other hand, we can see by looking at the price, price per gigabyte, and price per IOPS that both the Mtron and the MemoRight are significantly more expensive than a premium Disk Drive, being 2x to 3x more expensive, though not dissimilar on a per IOPS basis.

Similarly, if we consider the MLC drives by themselves, we see that these are a lot less expensive - often at only a third of the cost per drive and per gigabyte. But we also see that the IOPS performance is not all that good - effectively a little bit slower than a 7200 rpm drive, and about 1/5th to 1/10th of the Mtron and MemoRights. We see this reflected in the cost per IOP where the MLC drives are typically around twice the IOPS cost of a well founded SLC drive.

A Note About the IOPS-Gig Measure

Earlier, we expressed two common measures of assessing the cost or value of products: the cost per gigabyte and the cost per input-outputs per second (IOPS). One extremely useful measure to assess the relative value of a product combination is what we call the IOPS-Gig. This divides the total cost by the number of usable gigabytes and then further divides the result by the number of IOPS per device. The resulting value is expressed in mils (tenths of a cent). Thus, in the case of a 15k SAS drive, 39 mils times 250 IOPS a second times 36gb of space results in 351,000 mils, or $351.

The value of IOPS-Gigs is that they give you a general appreciation of the overall value of the solution, what ever its technical advantages vis-a-vis a given mix of IOPS and space. We see this in the comparison between the 15k SAS drive and the 7200 rpm drive. We see there is an approximate difference of 3.5:1 in IOPS-Gigs, and we realize that we do not assess these products purely by their relative cost which is 13:1 in terms of cost per gigabyte, and rather apply some sort of qualitative modifier (IOPS) to both prices.

You will note in this that while bare Flash Drives are not hugely dissimilar in IOPS-gigs, both are dramatically different from the IOPS-gig results obtained with MFT

Cost and Performance Conclusions for Various Flash SSD Models by Themselves

Model Capacity
in
GB
Drive
Market
Price
MS
Exchange
IOPS
Price
per
gig
IOP
Price

Mils
per
IOPS-Gig

2,5" 15,000 rpm SAS Hard Disk 36 350
250
9.72

1.40

39
2.5" 7,200 rpm SATA Hard Disk 200 150
70
0.75 0.75 11
Mtron Mobi 32gb 32 643 394 20.09 1.63 51
Mtron Mobi 64gb 3.5" 64 1,216 394 19.00 3.08 48
Mtron Pro 32gb 32 805 395 25.16 2.04 64
Mtron Pro 64gb 64 1,517 395 23.70 3.84 60
MemoRight 32gb 32 879 763 27.47 1.15 36
MemoRight gt 64gb 64 1,989 771 31.08 2.58 40
MemoRight gt 128gb 128 3,539 771 27.65 4.59 36
SuperTalent 32gb MLC 30 299

195

9.97 1.53 51
SuperTalent 64gb MLC 60 479 195 7.98 2.46 41
SuperTalent 128gb MLC 120 669 195 5.42 3.43 29
Mtron Mobi-1000 32gb MLC 32 290 83 9.06 3.51 110
Mtron Mobi-1000 64gb MLC 64 476 83 7.44 5.76 90
MLC CF 32gb kit 32 179 11 5.59 15.84 495
Proprietary 32gb MLC 32 179 50 5.59 3.61 113
Addonics PCI MLC CF 64gb kit 64 329 11 5.14 29.11 455

The Superior Performance of an MFT Enhanced Drive

The next table shows the same basic data as before, but reflects both the additional cost of adding MFT software, and on the other hand the dramatic increase in IOPS that MFT provides. This increase in speed is most apparent in the Mtron Flash media, because Mtron is well renowned for its extremely high random read speed, which complements the high random write performance of MFT.

The first thing to note, even from the first pricing element, is the dramatic improvement in performance with MFT. The quite good performance of the Mtron jumps almost 20-fold, to a point where it is 30 times faster than the fastest hard disk made. Similarly, MFT enabled drives are 10 times faster than the fastest current technology - the MemoRights.

But the most important thing to note regarding the Mtron and MemoRight items is that this has not been done at high cost. In general, the cost per IOPS of the SLC drives is barely 1/10th of the previous bare-drive IOPS costs.

The most notable general point of the table is the vast improvement in the performance of MLC drives. These incur a speed improvement of over 50-fold, and in the case of drives made from CF cards, almost 100-fold. As a result, the cost per IOPS of these falls below that of SLC drives, while the price per gigabyte of these is half that of an SLC drive. This is important to remember because the base cost of MLC media is expected to decline by at least two-thirds over the next three years.

A Note About MFT Pricing

You will note that there are two different base prices here for MFT. MFT has a pricing level for MLC drives that is half that of SLC drives. Similarly, the server variant of MFT is priced per 32gb of net usable storage space. MFT does not charge for parity or mirror drives. This is reflected in the price shown, which reflects the assumption that MFT will only get used in about three out of four drives of a drive array. The actual price of MFT may be lower based upon the storage array design, as well as the vendor providing product.

Cost and Performance Conclusions for Various MFT Enabled Flash SSD Models
Used in a Server Environment

Model Capacity
in
GB
Drive
Market
Price

MFT
Allowed
Cost

MS
Exchange
IOPS
Price
per
gig

IOP
Price

Mils
per
IOPS-Gig
2.5" 15,000 rpm SAS Hard Disk 36 350 n/a 250 9.72 2.58 39
2.5" 7,200 rpm SATA Hard Disk 200 150 n/a 700 0.75 1.20 11
Mtron Mobi 32gb 32 643

375

7,341 35.35 0.14 5
Mtron Mobi 64gb 3.5" 64 1,216 750 7,341 34.13 0.27 5
Mtron Pro 32gb 32 805 375 7,931 40.97 0.15 5
Mtron Pro 64gb 64 1,517 750 7,931 39.36 0.29 5
MemoRight 32gb 32 879 375 5,648 43.54 0.22 8
MemoRight gt 64gb 64 1,989 750 6,281 47.55 0.44 8
MemoRight gt 128gb 128 3,539 1,500 6,281 43.74 0.80 7
SuperTalent 32gb MLC 30 299 187 4,705 17.58 0.10 4
SuperTalent 64gb MLC 60 479 375 4,705 15.38 0.18 3
SuperTalent 128gb MLC 120 669 750 4,705 12.53 0.29 3
Mtron Mobi-1000 32gb MLC 32 290 187 5,463 16.58 0.09 3
Mtron Mobi-1000 64gb MLC 64 476 375 5,463 14.77 0.16 3
MLC CF 32gb kit 32 179 187 2,116 12.73 0.17 6
Proprietary 32gb MLC 32 179 187 2,166 12.73 0.17 6
Addonics PCI MLC CF 64gb kit 64 329 375 2,116 12.22 0.33 6

MFT for Laptops and Workstations

The workstation variant of MFT is much less expensive than the server version. The retail price of MFT for drives of any size used in Laptops or Workstations is just $149. The price shown reflects pricing of Internet sellers for such products. Here, while we see the dramatic performance gains of MFT, we also see the same reflected in prices that are very interesting.

Cost and Performance Conclusions for Various MFT Enabled Flash SSD Models
Used in a Laptop or Workstation Environment

Model Capacity in
GB
Drive
Market
Price

MFT
Allowed
Cost

MS
Exchange
IOPS
Price
per
gig

IOP
Price

Mils
per
IOPS-Gig
2.5" 15,000 rpm SAS Hard Disk 36 350 n/a 250 9.72 1.70 39
2.5" 7,200 rpm SATA Hard Disk 200 150 n/a 700 0.75 0.34 11
Mtron Mobi 32gb 32 643

129

7,341 24.62 0.11 3
Mtron Mobi 64gb 3.5" 64 1,216

129

7,341 21.44 0.09 3
Mtron Pro 32gb 32 805

129

7,931 29.78 0.12 4
Mtron Pro 64gb 64 1,517

129

7,931 26.24 0.10 3
MemoRight 32gb 32 879

129

5,648 32.14 0.18 6
MemoRight gt 64gb 64 1,989

129

6,281 33.77 0.17 5
MemoRight gt 128gb 128 3,539

129

6,281 29.24 0.15 5
SuperTalent 32gb MLC 30 299

129

4,705 14.56 0.10 3
SuperTalent 64gb MLC 60 479

129

4,705 10.34 0.07 2
SuperTalent 128gb MLC 120 669

129

4,705 6.62 0.04 1
Mtron Mobi-1000 32gb MLC 32 290

129

5,463 13.36 0.08 2
Mtron Mobi-1000 64gb MLC 64 476

129

5,463 9.65 0.06 2
MLC CF 32gb kit 32 179

129

2,116 9.82 0.15 5
Proprietary 32gb MLC 32 179

129

2,166 9.82 0.14 5
Addonics PCI MLC CF 64gb kit 64 329

129

2,116 7.30 0.11 3

MFT versus Zeus IOPS

As was noted at the beginning, the Zeus drive, by equivalent calculation, performs random IO of the size specified at about twice the speed of the MFT when used with SLC media, but a cost typically ten times as high. This is due in significant part to Zeus's transfer time which is 220mb/sec linear read and 115mb/sec linear write, as well as a random access time reported to be 20 to 100 microseconds, with a presumed average of around 25 microseconds.

Two points should be made about this. The first is that there are absolute limits to what any server can accept. The fastest time achieved so far has been 71,000 8kb random reads, or 541mb/sec on a 21 drive array. This is only 39% of the theoretical maximum for the same number of Mtron Mobi drives. It points out the obvious that computers and disk controllers have finite limits regarding the data they will accept. By comparison, our own tests showed that four drives could reach a rate of more than 30,000 8k reads, while 8 drives could exceed 40,000 8k reads. The practical point here is that there is only so much data that can be pumped into a Buss. Similarly, there is only so much data that can be pumped through an iSCSI, Fibre-Channel, or Infiniband network. Whether that point is reached with six or seven Zeus products or a larger number of something else, there is a point that cannot be exceeded.

Similarly, the size of needed storage should be considered. One can attain speed either through single big elements, or several smaller elements. But if one needs large amounts of storage as well as speed, paying for absolute speed rather than incremental speed makes little sense. For instance, if one needs two terabytes, one must practically engineer with 16 128gb drives, and whether these drives operate at 8,000 8k random reads a second or 4,000 random reads a second has little practical difference as both will hit the throat limit of 70,000 transactions a second.