The SSD myth unraveled

[Harm Millaard]

I wrote an article on SSD's, very short but maybe it helps to cut through all the myths of SSD's that infect all kinds of fora, including here.

The SSD myth unraveled

Written September 2011.

Introduction

SSD's are hot. Everybody talks about them and they are rumored to be the non-plus-ultra for performance gains. I write this to create a more realistic view on where we are now with SSD's.

SSD's have the reputation of being very fast, much faster than conventional hard disks. There are many reports on the internet that investing in SSD's will give you huge performance benefits and that is the explanation why so many people consider SSD's a must have for the ultimate performance experience. Unfortunately - and this is especially true for video editors - this is mostly a hype and not a wise decision.

Let's start with the basics.

They are physically small, have no moving parts, are quiet, cool and expensive per GB. The small physical dimensions mean that you can easily fit 4 SSD's in a single 5.25" bay. Because there are no moving parts they are quieter than conventional disks. They also operate at lower temperatures than conventional disks, which is a distinct advantage in a crowded system. Average access time is negligent in comparison to conventional disks. SSD's are not as susceptible to fill-rate degradation as conventional disks. (Conventional disks tend to show performance degradation when filled for more than 60%. Not so for SSD's). But there is the cost aspect and the reliability question.

But the most important question is performance. Does it justify the extra cost for the increase in performance?

Currently, the price per GB for a SSD of the latest generation is generally around $ 1.40 - $ 2.00, depending on the model, capacity and brand. A conventional disk is around $ 0.05 - $ 0.06 per GB and that means a SSD is around 30 times more expensive per GB. Is it worth the difference?

According to many, the answer is yes, it is worth it, but I beg to differ. Proponents of SSD's claim that the transfer rates of SSD's with figures of 500 MB/s are way faster than the 120+ MB/s of modern conventional hard disks and that justifies the extra cost. If this were true, why don't we see those performance gains in our benchmark? What is wrong with these claims of unprecedented speeds?

The background

Manufacturer's claims of IOPS - which are irrelevant for video editing - and sequential transfer rates are based on highly compressible data in 4K blocks, something that video data are not, because they have already been heavily compressed. It boils down to writing only 0's and compressing those 0's to achieve the claimed transfer rates of 500 MB/s, but effectively only around 30 MB/s are transferred. If you were to test effective transfer rates using CrystalDiskMark, which uses random data to benchmark, the compression is far less, because it is random, and then the effective transfer rate of SSD's is reduced to something in the order of 200 MB/s. With video data, the effective transfer rate could well be even much less because of the heavy compression that has already taken place during the shoot.

Interesting to see, and I do not know the answer, whether a heavily compressed codec like AVCHD would show lower transfer rates than less compressed codecs like P2-Intra or 50 Mbps MPEG2.

Write degradation

This is one of the most discussed issues with SSD's. On new SSD's the write speed is almost as good as the read speed, but when using that SSD for a longer time there are serious performance issues while writing data to a SSD. Even with the latest generation SATA3 / Sandhurst SF-2281 SSD's, write performance can easily drop by more than 60% despite the TRIM function. This effectively means in the best case scenario, that a SSD with a claimed transfer rate of 500 MB/s, which delivers less than 200 MB/s read speed with video data, can only deliver 80 MB/s or less write speed when used for some time. That is not too impressive in comparison to conventional disks at a fraction of the cost and less than a simple raid0 with two conventional disks on a ICHR10 on-board controller attains.

If TRIM is not working, the write degradation is even worse and you may count yourself lucky to attain write speeds of 50 MB/s or less. Unfortunately, most SSD's firmware in combination with raid controllers currently have the nasty side effect of disabling the TRIM function, so raiding SSD's is not a serious choice for raid configurations.

The only way to correct this write degradation is by performing a secure erase, which means losing all the data on the SSD, not a nice perspective for anybody, but most of all for notebook users. Are your backups current?

Reliability

NAND memory is susceptible to ageing and most SSD's calculate their lifespan in data transfer values. In a worst case scenario this usually means you can rewrite the complete contents of a SSD around 125 times, before the NAND memory is no longer reliable/useable. Not many people would try that and for a boot disk this means a very, very long time before the useful life of a SSD is at an end, but for video editors it is a different story.

The bottom line on SSD's at this moment of writing

They are the way to go in the future, but not yet.

For OS & program disks they are great, provided you set up Windows to not use the SSD for temp storage. They can easily shave off 3 or 4 seconds from your usual boot time of 60+ seconds, depending on your configuration. (Did you notice any sarcasm in this statement? You should.) They are a waste of money for video storage and do not deliver any performance gain, because of the compression that has already taken place with the video material and that lowers the transfer rates significantly. The faster loading of programs, which is often used as an argument for SSD's, is usually limited to 1 second per program or not even be noticeable.

Question

What would you rather have at this moment:

a. 8 TB of net storage with conventional disks for around $ 400 with a sustained transfer rate of 1000 MB/s, or
b. 2 TB of net storage with SSD's for around $ 3000 with the same transfer rate?

Just my $ 0.02

[Chris Hurd]

Thanks -- offsite link replaced with full text of article.

Please help keep eyeballs on DV Info Net. To share your article here, just post the full text of the article itself instead of a link to another forum. Much appreciated,

[Mike Peter Reed]

"They can easily shave off 3 or 4 seconds from your usual boot time of 60+ seconds, depending on your configuration. (Did you notice any sarcasm in this statement? You should.) They are a waste of money for video storage and do not deliver any performance gain, because of the compression that has already taken place with the video material and that lowers the transfer rates significantly. The faster loading of programs, which is often used as an argument for SSD's, is usually limited to 1 second per program or not even be noticeable."

Not my experience at all (early 2009 iMac - internal HDD failed so booting Lion from SSD over FW800)

Boot time - halved (conservative estimate)

App launch time - quartered (conservative estimate), sometimes "instant"

More expensive? Yes

The future? Yes

The now? FUD

The current legacy? spindles

I agree that it's only possible to write to each sector on a SSD millions of times before expected failure.

Note - I am only using SSD because the internal HDD failed after 18 months - MTBF for HDD is generally rated much longer of course.

As ever, measure the cost to benefit ratio and choose your poison.

[Chris Hurd]

In my direct experience, operating systems (both PC and Mac) boot considerably faster from SSD's than from hard drives.

[Harm Millaard]

That is correct, it will save you 3 or 4 seconds, but the whole boot sequence with the POST tests, the spin-up of your drives, the initialization of your raid controllers, the loading of the controller BIOS, etc. still take the same time, so the first 60+ seconds are spent anyway. It is only after that initialization that a SSD can kick in and then it saves you - normally once per day - 3 to 4 seconds. So instead of 70 seconds for booting if will only take 67 seconds or there about. Does that seem like a good argument to tell your wife why you need to spend that much on your new system?

As to loading your applications, that allegedly happens much faster, I can not discern any difference. First time loading of PR CS5.5 with conventional disks happens in around 7 seconds and second time in around 3 seconds. No discernible difference with SSD's in my case, but even if there were a 1 second difference, does that justify the cost difference? Not IMO.

[Rob Wood]

thx for posting.

[Bart Walczak]

While I agree with general idea that Harm presented, I'd like to point out that 120MB/s+ is the speed of the newest generation of 7200rpm hard drives. If you are still using for example older hdd that has 30MB/s or 50MB/s, then you will see more difference when upgrading to SSD, however you will also see the difference when buying a new HDD.

In case of notebooks, of which many still use 5400 rpm drives, the use of SSD has more sense due to much better performance, smaller form factor, and also power consumption.

Personally however I think that the reliability is not there yet, and the pricing is still ridiculous. The only place where the use of SSDs is justified is in high-throughput web and database servers, where IO number is often really what determines the performance. Video editing - not so much.

[Steve Kalle]

I can see someone doesn't like change.

There might be some hype about SSDs but a lot of it is TRUE.

Your premise that people are looking to use SSDs as storage is WAYYYY off. I haven't met or spoke to a single person who wants to use SSDs for storage. Everyone considering SSDs want to use them for the OS.

Most people looking for a 'performance' drive for the OS previously sought after the Raptor/V-Raptor. Looking at today's prices of a 150GB V-Raptor at $130 and a 120GB OCZ Vertex at $140, WHY WOULD YOU CHOOSE A VELOCIRAPTOR? With the OS and Programs, mechanical disks cannot come close to the performance of 95% of current SSDs

In terms of reliability, some of the largest web servers have released the annual rate of failure. Guess what? SSDs have a FAR lower failure rate than the best 15k Enterprise dirves.

WRITE DEGRADATION: Totally something of the past. Every manufacturer worth anything has implemented a program within the firmware to fix this issue. Intel was the first and still is the best. The only time you will see a significant reduction in write performance is when you constantly fill the entire drive using certain programs like HDTach or HDTune, and fill them over and over immediately after each completion.

To your point about this 60 seconds, not everyone has multiple pieces of hardware and software being loaded at boot up. Also, I consider 'total boot up time' to end once you can actually begin clicking and opening things such as programs or the internet. Just saying boot up ends when you see the desktop background is not looking at the entire picture. SSDs cut down even more time when you include this.

I just ran my own test of a V-Raptor vs an Intel X25 SSD with PPro: 19 seconds on the VR and only 7 seconds on the SSD. An unrealistic test would be to close PPro and then re-open it because no one does that. Also, closing and opening allows Windows to store parts in ram and/or cache in order to make loading faster, but who closes Premiere and opens it for the heck of it? Normally, you open it, do you work and then close it; then come open it later in the day or the next day. By then, nothing is pre-loaded (maybe its called pre-fetch in Windows).

I just did another test clicking on AE, PPro & PS and everything was loaded in 8 seconds. Try doing that with ANY mechanical disk within the next 20 years.

Anyone doing panoramic work and/or large megapixel images in Photoshop, using a SSD as cache massively helps.

For fun, I just opened AE, PPro, PS, Cinema 4D, MS Word, Excel and Sony ClipBrowser in 10 seconds on my SSD. (essentially everything I have stored on my taskbar)

I highly suggest anyone wanting to know the truth go to Anandtech and Tomshardware and read their lengthy papers. One paper I read was 34 pages long which helped me understand SSDs. I was like Harm before reading these papers - I didn't want anything to do with SSDs. After reading these papers, I bought an Intel X25 within a week.

While I don't think SSDs are economical as storage, they do have one very good feature for storage purposes: enterprise level Unrecoverable Bit Error Rate of 1x10^16 (at least, the Intel 320 SSD does). Compare this to conventional disks with a UEBR of 1x10^14. Every 12TBs of reads, there is an unrecoverable error which can cause file corruption or 'in the worst case scenario', can cause an entire raid array to be lost during an expansion or rebuild. With 1x10^16, it takes 1250TBs of reads.

[Harm Millaard]

Your premise that people are looking to use SSDs as storage is WAYYYY off. I haven't met or spoke to a single person who wants to use SSDs for storage. Everyone considering SSDs want to use them for the OS.

You will be surprised. Currently with 449 systems in the PPBM5 benchmark, there are 141 desktops and 7 laptops with a SSD for OS. Out of those, there are 37 desktops and 3 laptops that also use one or more separate SSD's for storage and projects.

Your statement about Velociraptors is antiquated and belongs in the historic museum. That was accurate more than two years ago, but the technology has improved so much that even a $ 40 SATA disk is much faster than a 150 GB Velociraptor.

Write degradation still happens with all the latest generation SSD's with the latest Sandhurst controllers when you are editing and render your timeline from time to time after modifications. And significantly so. It is not only with applications like HDTune, but during normal editing and rendering the timeline.

I just ran my own test of a V-Raptor vs an Intel X25 SSD with PPro: 19 seconds on the VR and only 7 seconds on the SSD.

That only tells that your system is not properly configured or there is something wrong with your Velociraptor, because I can repeatedly get 7 seconds on the Velociraptor, including several third party plug-ins and I know from SSD users that can not get faster than 8 seconds on SSD. Same with all the other applications from the Master Collection, AU, EN, DW, PS, AE, they all start within 7 seconds or less. AU starts in around 3 seconds and EN in about 5 seconds.

[Pete Cofrancesco]

You are off on the boot time. I went from 1.5 minutes on my old system to 10-12 seconds with SSD and 30-40 seconds with HDD on a new system. Boot time just tells part of the story, launching apps, virus scans, etc, are lighting fast. The power consumption, heat generation and noise savings are things I notice and enjoy.

Traditional hard drive have run up against physical speed limitations of their technology long ago. Drives have gotten bigger and cheaper but not faster. In some cases 2 tb drive actually have gotten slower and less reliable. SSD numbers are written to favor the manufacture's sales but they have always been for all technology case in point hard drive size or usb/sata speed etc. Cut the numbers in half and SSD is still twice as fast. Numbers are numbers, in my real world experience SSD performance benefits are unmistakable and that's all that matters to me.

Per megabyte sure SSD is expensive but when you look at the speed gains you are getting its actual a good value. What's an extra $100 when building a new system? Agreed in video its not any magic bullet but who ever said it was. Sounds like you've made up your mind against SSD, that's fine. When you want to argue about the worth of something everyone is going to have there subjective option.

[Randall Leong]

Originally Posted by Steve Kalle

I just ran my own test of a V-Raptor vs an Intel X25 SSD with PPro: 19 seconds on the VR and only 7 seconds on the SSD.

I agree with Harm on this: Your particular V-Raptor is either faulty or more than two-thirds full. Any hard drive that's more than two-thirds full will slow down significantly - both in sequential transfers and random accesses.

I did my own testing on an i3-2100 with 4GB of RAM and a WD Blue 500GB drive launching PPro CS5.51. I came up with about 7.5 seconds from the time that I clicked on the Premiere Pro link on the Start menu to the time that the window to create a new project or open saved project came up. But then again, that hard drive was only about 10% full when I tested. What's more, that particular WD Blue was slower in random accesses than all of my other 7200 RPM hard drives. In the end, the 15 ms difference in random access speeds between the WD Blue and a typical SSD results in a less than 1 second difference in PPro launch performance.

[Daniel Browning]

Harm, I didn't see anywhere in your post where you dealt with concurrent sequential I/O, that is performance while reading multiple video streams simultaneously. A single concurrent sequential I/O op only reflects video editing with a single video track and no dissolves. Personally, I will frequently have 4-5 layered video tracks at some points in the timeline. I can easily hit 300 MB/s with my 5-disk RAID-5 and read/write I/O with my 8-disk RAID-10, but only if I'm doing one video stream at a time. When I start reading 2, 3, or more, the I/O degrades nonlinearly (i.e. the sum of all transfers is less than 300 Mb/s). With SSD, I get a perfectly linear degredation. 20 streams at 10 MB/s each or 1 stream at 200 MB/s.

[Harm Millaard]

Daniel,

How do you measure that? Because then I can check on my side what happens with multiple flows.

[Daniel Browning]

Originally Posted by Harm Millaard

How do you measure that?

Sorry, I forgot what it was called. It was a Windows GUI program that let you dial in the number of simultaneous read streams for the test, as well as a huge number of other parameters. It ran the entire test several times (something like 10 by default, I changed it to 3 because I didn't want to wait a week for the results.) and wrote the results to a text file as well as displayed on the screen. I tried googling it just now, but I was in a maze of twisty benchmark programs, all alike.

Originally Posted by Harm Millaard

Because then I can check on my side what happens with multiple flows.

That would be neat -- sorry I'm about as helpful as a box of rocks. :)

[Daniel Browning]

Originally Posted by Daniel Browning

Sorry, I forgot what it was called.

Duh, I googled the internet but I forgot to check my own computer. I found the program. It's called "iometer":

Iometer project

Not the most amazing website I've ever seen. :)