About 50GB a day writes, how long would a 256 840 pro last?

[Hugo Drax]

Trying to figure out the replacement cycle/cost. Is there some kind of SSD expected life calculator where you can plug in numbers and get life expectancy numbers?

 

[Coup27]

There are programs such as SSDLife and HD Sentinel to name a few which will report an estimate of your remaining life. This article may be of interest to you.

http://www.anandtech.com/show/6459/samsung-ssd-840-testing-the-endurance-of-tlc-nand

With an estimated write amplification of x3 and no additional spare area, around 14 years would be best you could calculate without working out your write amplification.

 

[Lonyo]

http://www.anandtech.com/show/6459/samsung-ssd-840-testing-the-endurance-of-tlc-nand

SSD Lifetime Estimation
NAND 	               MLC—3K P/E Cycles              	TLC—1K P/E Cycles
NAND Capacity 	       128GiB 	256GiB 	          128GiB 	256GiB
Writes per Day 	        10GiB 	10GiB 	           10GiB 	10GiB
Write Amp               3x 	3x 	                   3x 	3x
Total Estimated Lifespan 35.0 years 70.1 years  11.7 years 	23.4 years

About 23.4/5 years, or 4.68 years.

 

[birthdaymonkey]

Is there a reason that AT article assumes a WA of x3? Is that a worst case scenario? I'm sure I've seen a comparison of modern drives that indicates most are fairly close to 1 in normal use.

 

[Coup27]

About 23.4/5 years, or 4.68 years.

The 840 Pro uses MLC NAND.

 

[Lonyo]

The 840 Pro uses MLC NAND.

My bad, 70/5 or 14 years.
Missed the Pro in the title.
Information was still in the same article/table though 😛


For Write Amplification:

Write amplification can also go over 10x if your workload is heavily random write centric, but that is more common in the enterprise side - client workloads are usually much lighter.

So you can just divide by 3.333 to get worst case of 10 WA, or about 4.2 years.

 

[Hellhammer]

Is there a reason that AT article assumes a WA of x3? Is that a worst case scenario? I'm sure I've seen a comparison of modern drives that indicates most are fairly close to 1 in normal use.

It's just an assumption with some wiggle room. Worst case WA can be much higher (and depends on how full and fragmented the drive is) but I think ~3 is a realistic, yet conservative figure for consumer workloads. It's somewhat based on WAs we have measured (hard since most drives don't report NAND writes).

 

[Hugo Drax]

Sounds good, my design criteria for the box was 3 years. so I should be good with the 30-50GB writes a day.

 

[Ao1]

A WA factor of 3 sounds very high to me but I guess it depends on the controller/ firmware. SF drives have low write amplification due to the compression advantage for tasks like installations, but conversely for just about any other task it is possible to see very high WA.

There are so many variables on hardware & f/w without considering work load. Garbage collection aggression, OCZ drives switching from SLC to MLC mode, large DRAM caches etc.

I found that 4K random writes do not burn out the media wear out indicator any quicker than sequential writes. Sounds crazy but obviously the write speed for 4K random writes is quite slow, whilst sequential write speeds are very fast.

Intel were touting 1.1 WA with their X25-M drives and it’s hard to see how that could ever be improved without compression.

 

[Cerb]

Is there a reason that AT article assumes a WA of x3? Is that a worst case scenario? I'm sure I've seen a comparison of modern drives that indicates most are fairly close to 1 in normal use.

You want some padding. I know CAD-type programs can get WAs near 2, and with consistently small writes, and there's not much any fancy WA-reduciing features (like compression/dedup) can do. The buffers need flushing, and there's simply not much data to flush at a time...but, it comes out GBs over the course of a day. I'm sure worse can be found, without adding OLTP DBMSes. Worst-case will be much higher, but that should also be very rare, and typically balanced out by many times more that falls within expected (read: optimized for) workloads.

3 should be higher that what you'll see, but you don't want to assume 1-1.5, and then find you're risking bitrot before your planned upgrades, from actually having much higher.

For most of our uses, anything with with MLC will last long enough that it will be obscenely obsolete by the time it gets chewed up.

 

[GAO]

With light use on my home PC, I have written 1 TB on my 128GB 840 Pro with a wear count of 13, so that is about a 1.66 write amplification. FWIW.

 

[Ao1]

You want some padding. I know CAD-type programs can get WAs near 2, and with consistently small writes, and there's not much any fancy WA-reduciing features (like compression/dedup) can do. The buffers need flushing, and there's simply not much data to flush at a time...but, it comes out GBs over the course of a day.

That is a good point. Window’s also issues a steady stream of small writes, which I suspect is the worst write pattern for WA as the buffer has to be periodically flushed resulting in write patterns that are not optimised to reduce wear. Below are some stats taken over 4 hours with little activity outside of OS background tasks. Some of those stats look like they are based on a HDD, but they were taken using an 840 Pro and are fairly consistent of light OS use. (Win 8 64Bit, Stats taken from Physical Device)

· Number of I/0’s performed: 23,974
· Amount of data xfered: 604.57 MiB
· Percentage of I/0 operations that performed a random access: 92.97%
· Percentage of data xfered via random I/0 operations: 70.32%
· Minimum response time: 0.021 ms
· Average response time: 5.756ms
· Max response time: 984.644 ms
· Percentage of Fast I/0’s: 92.1457%
· Max IOPS: 722
· Percentage of I/O operations above queue depth: 1 17.47%
· Max MBs range: 10 < 20 MBs
· 271.45 MiB xfered within response times brackets: 1 ms < 5 ms
· Number of I/0&#8217;s with a response time more than 500ms: 24
· Number of I/0 operations below 4 KiB: 601
· Percentage of 4 KiB xfers that were random: 94.57%
· Percentage of I/0&#8217;s using 4 KiB data xfer size: 56.75%



Uploaded with ImageShack.us

 

[SSBrain]

If it can be of any interest, at 1.875 TB of total host writes, write cycles on my Samsung 840 250 GB (TLC NAND) increased to 9. Taking into account 256 GiB of flash mememory in total (of which 250GB -> 232.83 GiB are available to the user), that implies a write amplification of 1.2 in my case. Most Samsung 840/840 Pro owners I've talked to, with home usage (which includes frequent system shutdowns) have a write amplification of around 2, though.

It appears that the more the drive idles, the less the write amplification gets. Probably this gives the SSD controller's garbage collector more time to optimize its flash memory usage. I rarely shut down my systems, which it's probably why my SSDs' WA is lower than usual (not only this one, but also a couple other Samsung 830 SSDs in different PCs used in similar conditions)

Of course, actual data workload also affect write amplification. With home workloads it shouldn't get very high, though.

To answer the OP's question, assuming a write amplification of 2 (average consumer workloads), I would expect a minimum total life of at least:

[256 GiB] * [3000 P/E cycles] / [2 WA] = 384000 GiB of writes

At a sustained average of 50 GiB/day (which is *a lot*), it would mean that the drive would last 21 years.