System Testing and Benchmarks

[NoelC]

Impressive!  Especially so given the 512GB can be had right now for $219 at Newegg.com and Amazon.com.

 

For those without SSDs it's hard to find an excuse not to upgrade at that price. 

 

Such an upgrade is likely the single biggest thing you could do to make your system hugely more responsive for just a few bucks.

 

Do you know of anyone who's set up two or four of them in a RAID array yet?

 

-Noel

Edited June 18, 2014 by NoelC

[TELVM]

2x MX100 256GB RAID0:

 

 

^ Notice that 4K random QD1 speeds don't improve and even drop a bit in RAID0 ...

Edited June 18, 2014 by TELVM

[TELVM]

One of the latest iterations on the "fast consumer SATA III SSD" theme, the Sandisk Extreme Pro:

 

 

http://www.thessdreview.com/our-reviews/sandisk-extreme-pro-ssd-review/4/

 

(^ Again the "sonic wall" at ~36MB/s in 4K random QD1 read)

 

These things cost almost double than a MX100, while in the real consumer world we'll be hard pressed to notice any difference in speed between the two models.

 

 

 

 

- ADDENDVM - Very interesting ongoing experiment about real world SSD longevity:

 

The SSD Endurance Experiment: Casualties on the way to a petabyte

Edited June 19, 2014 by TELVM

[jaclaz]

- ADDENDVM - Very interesting ongoing experiment about real world SSD longevity:

 

The SSD Endurance Experiment: Casualties on the way to a petabyte

Very, very nice!

 

jaclaz

[NoelC]

I enjoyed that article on longevity very much, thank you.

 

I would suggest not focusing quite so hard on the 4K QD1 numbers.

 

Single threaded 4K writes and reads by definition can't exceed the speed of a single drive - a data block that small doesn't get broken up across multiple drives in the array, so basically the process goes like this:  Queue a 4K block write, RAID controller selects a single drive to receive the data, drive writes, drive signals it's done, software is notified.

 

Good thing in the real world data blocks much larger than 4K are typical, I/Os are queued deeper than 1, we have caching, and we have multi-threaded systems.

 

-Noel

[TELVM]

... I would suggest not focusing quite so hard on the 4K QD1 numbers ...

 

I happen to cordially disagree    , and would dare to say: Don't focus so hard on the mesmerizing sequential numbers.

 

 

" ... Fast sequential speeds allow for quick file copies and smoother performance when working with large files, like videos. However, it is random performance, measured in Input/Output Operations Per Second (IOPS) that is, perhaps, the most important performance metric for SSDs. A large portion of storage activity is made up of 4K random writes, a metric that measures how well a drive will perform when writing small chunks of random data (e.g. changing a small piece of a Word or text file and then saving the changes). Users spend a majority of their time not copying large files or installing applications, but multitasking (e.g. email, web-surfing, listening to music, etc.) and working with various work and media files - tasks influenced by IOPS ..."

 

Why SSDs Are Awesome - An SSD Primer

[NoelC]

In all seriousness, I look most at the relatively small I/O performance at higher queue depths.  And at the small but not tiny sequential operations.

 

These are the things that affect day to day desktop life the most.  How fast you can open your application (such as Photoshop, and to subsequently open an image that's a few megabytes), how quickly you can compile your Visual Studio project, how quickly you can display your menus.

 

You have a basic minimum latency time with a modern computer, measured in the microseconds, to access the I/O subsystem because it's just that - a subsystem that is not closely integrated with the CPU.  That's really not going to change fundamentally until such time as the chipset manufacturer (Intel) starts building terabytes of permanent storage into the processor's address space.

 

Not too awful long ago SSDs came along and reduced that latency from milliseconds to microseconds, and they did cause a fundamental change in responsiveness.

 

Have you worked with SSDs for long?  With arrays of them?  In my experience having BOTH decent 4K response (which any SSD gives) AND gargantuan sequential speed is important for all around performance, and having a bunch of cores to generate separate I/O streams doesn't hurt either.  For me Photoshop starts cold in 3 seconds and opens gigabyte sized files in just a few more.  IE and Word and Explorer all come up instantly.  My Visual Studio products (with hundreds of thousands of NCLOC) compile in 10 seconds.  With a system this responsive a new kind of fatigue can set in - one tends to multitask even more and there's never a time for a coffee break!  That said, I wouldn't mind testing an array of 4 x 512 GB MX100s, if someone would send me a set...

 

-Noel

[TELVM]

Samsung 840 EVO (single 250GB SATA-III SSD):
 

 

 
^ Those "supersonic" 43 MB/s @ 4K random read QD1, that's the stuff I like the most in a SSD     .

[NoelC]

I don't know where you got those benchmarks from, but I'd be willing to bet in large part that that enviable 4K speed is more because it's being measured on a really happenin' computer system, and less because an 840 EVO is somehow especially good by comparison to its peers.   

 

By all accounts an 840 Pro should be better. 

 

Note that the average access times are not as small as seen on other drives, for example.

 

-Noel

Edited June 25, 2014 by NoelC

[TELVM]

That benchmark was on a 4670K / Z87 Maximus VI Extreme, nothing to write home. But the guy who posted it knows a couple things about SSD optimisation.
 
 

"... Now, sequential speeds are not really what make SSDs so fast and desirable for the OS. It is their 4K read speeds and access times. 500+ sequential reads on SATA 3 is a nice upgrade from the 100MB/s on HDDs, but the 4K reads are usually 20-40 times faster than that of a hard drive! (~20MB/s vs .5MB/s!) The faster the 4K read speeds the faster your system will be ..."

The 840 EVO uses of a couple tricks (larger DRAM cache, "turbowrite") that in some situations give it the edge on the Pro.


(BTW Samsung is about to launch a new SSD model that sounds interesting, stay tuned ...)

Edited June 26, 2014 by TELVM

[NoelC]

Write-back RAM caching inside the drive...  Write-back RAM caching in the Intel RST driver...  Write-back RAM caching in the file system...  Where's the benefit of where it's done besides being able to game the benchmark programs? 

 

If you have your system set up well your file system RAM cache is already handling your I/Os anyway, and probably far better than a small cache inside the SSD.  Your software is not waiting.  In fact, having more data copy operations (to and from too many different RAM caches) can actually slow things down in a big picture throughput sense.

 

Do you have your file system Write Back Buffer Flushing disabled?

 

Because it can be affected by caching, this is a good example of where a specific reading in a specific benchmark result isn't all that representative of real-world results.  Again, I advise not focusing so hard on the single simple "issue a single I/O and wait for it to finish" process.  Real world operation is not that simple.

 

I've mentioned it before in this thread and posted screen grabs - try the Passmark Advanced Disk benchmark for Workstation in the PerformanceTest package, and set it to either the "Standard C/C++ API" or "Standard Win32 API (Cached)" setting if you want a VERY good measure of system responsiveness that actually gauges the performance you'll see in real use.

 

-Noel

Edited June 26, 2014 by NoelC

[TELVM]

There are obvious benefits from cache cascading, that's why some CPUs have level four caches nowadays. Think of a 840 EVO like a subordinated mini-computer for storage: ARM Cortex R4 < DRAM ("L1") < Turbowrite ("L2") < NAND.

 

Spinner makers are trying to delay the inevitable integrating SSDs for caching into hybrid HDDs, example.

 

Do you have your file system Write Back Buffer Flushing disabled?

 

Yep.

[NoelC]

I didn't say having multiple cache levels necessarily hurts things, but it can.  It's not always better to move data.

 

it's good to see the drive makers improving their controllers still, but they've hit the wall of not being able to reduce the latency below a few microseconds. 

 

Frankly the Intel approach seems to have more long term promise:  Eliminate the SATA link altogether and more closely integrate the controller with the CPU using an optimized protocol.  I imagine we'll be seeing some improvements on that front.

 

I'd love to know what's going on in the minds of designers of the chipsets at Intel.

 

-Noel

[TELVM]

...  (BTW Samsung is about to launch a new SSD model that sounds interesting, stay tuned ...)

 

Nope, nothing to write home after all:

 

Samsung 850 Pro review

 

 

 

Seems many SSD makers are just consolidating (lowered prices, extended warranties, etc.) more than pursuing breakthroughs in performance.

 

Until someone offers some SSD with 100 MB/s random read QD1 I see little reason to expend money beyond something like the MX100.

[NoelC]

 

SSD makers are just consolidating (lowered prices, extended warranties, etc.) more than pursuing breakthroughs in performance

 

My God, what a misrepresentation!  That's simply awesome performance for a SATA storage device!  Latency is 22 microseconds.  Nearly one tenth what it was just a couple of years ago.

 

You need to come to grips with the fact that the connection method (SATA III) is the limit here, not the drive.  You cannot get latency below zero!!!  There is a theoretical limit to the performance of a device connected via a SATA III link, because of things in the system, not the drive, and the modern drives are all getting close to the limit.

 

From the article:

 

 

Looking closely at the Samsung 850 Pro, we have a new release SSD that has topped every benchmark conducted, bettering the next best in our Vantage chart by almost 2K points. 

 

Again, I emphasize that these drives are splitting the hairs of difference between best in show and the theoretical maximum throughput based on today's SYSTEM technology.

 

-Noel

Edited July 2, 2014 by NoelC