SSD Write Endurance 25nm Vs 34nm

[johnw]

If I count only the 4K writes the WA is currently coming out at 29.47 (What I reported before included all write activity)

That is more like the WA I would expect to see for full-span 4KiB random write.

[Christopher]

So I'm entertaining ideas on what to do with this:



It's an MTRON PRO 7000 (msp7035). It is faster than I thought, but not by much. It's a 3.5" 16GB drive, and it's pretty old school.

But it's is brand new, never been opened.

EDIT
I ran ASU for a while, and after 70GB of writes, the one SMART attribute's RAW value went from 324393 to 453601, and was able to do about 51MB/s.

I'm going to have to take the drive apart as I'm curious to see whether the PCB inside is 3.5" or 2.5" inches among other things. It just seems ridiculous to put 16GB of NAND inside a 3.5" enclosure.

[Christopher]

Samsung 830 64GB Update, Day 33
FW:CXM01B1Q

GiB written:
205213.42

Avg MB/s
86.87

PE Cycles
13017, up from 12503 yesterday

Reallocated Sectors
40960
20 Blocks, up from 18

796 Hours





The 7 hours of idle time combined with halving the static data down to 18.5GB seems to have increased speed by 20MB/s. I'm not sure how long it will last.

[B.A.T]

Todays update.
Kingston V+100
327.5644 TiB
1801 hours
Avg speed 25.02MiB/s
AD still 1.
168= 1 (SATA PHY Error Count)
P/E?
MD5 OK.
Reallocated sectors : 00


Intel X25-M G1 80GB
216,5665 TiB
20109hours
Reallocated sectors : 00
MWI=142 to 141
MD5 =OK
50.41 MiB/s on avg


m4
199.1485 TiB
730 hours
Avg speed 80.31MiB/s.
AD gone from 243 to 240.
P/E 3502.
MD5 OK.
Reallocated sectors : 00

[Christopher]

MTRON claims a 32GB drive can write 50GB a day for 140 years, so assume 50GB a day for 70 years with the 16GB 7000.

That would equal 1,277,500GB, which is realistic if the NAND is rated for 100,000PE cycles. If the MTRON could maintain 3500GB per day in ASU, I could hit 1,277,500GB in only 365 days. 41MB/s avg would be 3500GB a day so that seems reasonable.

Of course, if the NAND is 200,000PE rated and the drive has decent WA/WL, then all bets are off. It could be more like 3 years. I only ran a couple of loops with no static data and 3GB min free space, so I can't say whether the lack of TRIM is going to wreck avg speed or not. But it could be feasible to test the drive in a reasonable time span -- maybe. The lack of SMART data is irritating (ONE attribute? Why bother?), and I don't even know how the one attribute works.

The 3.5" chassis is quite nice. It weighs almost as much as a single platter 3.5" HDD.

[minpayne]

According to the resource monitor the QD is 1.

Have you tried "Performance Monitor" of Windows 7? Sometimes I get strange readings from "Resource Monitor"...

[Ao1]

Average WA
60GiB (LBA) * P/E Cycles (308) = 18,480
Actual writes = 1,398
WA = 13.31

4K Random only WA
60GiB (LBA) * P/E Cycles (268) = 16,080
Actual writes = 666
WA = 24.14

MWI: 91
Avg write speed = 4.31MiB/s

In the Toolbox/ System Information it states that the MAX LBA is 125045424, so I am now using 60GiB rather than 64GiB.

177 is still only updating after a power cycle.

Have you tried "Performance Monitor" of Windows 7? Sometimes I get strange readings from "Resource Monitor"...

Yes and it reports the same.

[Ao1]

The 7 hours of idle time combined with halving the static data down to 18.5GB seems to have increased speed by 20MB/s. I'm not sure how long it will last.

Christopher, during the idle time did the SMART atribute 177 update?

[Anvil]

Kingston SSDNow 40GB (X25-V)

615.24TB Host writes
Reallocated sectors : 05 21
Available Reserved Space : E8 99
POH 5560
MD5 OK

33.38MiB/s on avg (~111 hours)

--

Corsair Force 3 120GB

01 89/50 (Raw read error rate)
05 2 (Retired Block count)
B1 70 (Wear range delta)
E6 100 (Life curve status)
E7 10 (SSD Life left)
E9 662250 (Raw writes) ->647TiB
F1 881407 (Host writes) ->861TiB

MD5 OK

106.81MiB/s on avg (~159 hours)

power on hours : 2533

WRD is all time high as far as I can see. (since the retention test it has been increasing steadily)

--

[minpayne]

Anvil, my opinion is that the "true" WA of Christopher's Samsung 830 64 GiB was around 3.93 under Endurance-testing of ASU, and the estimated "MWI exhausted" was supposed to be around 44.4 TiB if the P/E spec was 3000.

I'm also confused by your last figure, the "P/E Cycle summary". I suppose it's not the true P/E count, but instead, the P/E count divided by WA?

By the way, we can't wait to get the beta11 with the Enterprise-test

[Anvil]

The P/E Cycle summary is Based on

Host Writes / Capacity

I'll explain a bit more later, I haven't found a better title for the chart,yet.

[minpayne]

Then that's the same as what I thought:

Host Writes / Capacity == P/E Cycles / WA

I would call it "Reduced P/E Cycles after WA"

[Christopher]

Christopher, during the idle time did the SMART atribute 177 update?

Negative. Not even by one as it did in the beginning.

[Ao1]

Thanks for confirming that Christopher.

I just used hIOmon to see if TRIM operations had time to execute. TRIM with the normal ASU client workload works OK, in fact TRIM operations are lightning fast compared to SF. It took only 8.79ms for a 50GB file delete. SF = 20 seconds plus)
With the ASU enterprise workload however it seems that TRIM operations don’t occur and I'd guess that would be the same for any SSD with that workload.

[Christopher]

Samsung 830 64GB Update, Day 34
FW:CXM01B1Q

GiB written:
211183.25

Avg MB/s
70.25

PE Cycles
13321, up from 13017 yesterday

Reallocated Sectors
40960
20 Blocks, up from 18

820 Hours

[Christopher]

Double post

[Christopher]

Thanks for confirming that Christopher.

I just used hIOmon to see if TRIM operations had time to execute. TRIM with the normal ASU client workload works OK, in fact TRIM operations are lightning fast compared to SF. It took only 8.79ms for a 50GB file delete. SF = 20 seconds plus)
With the ASU enterprise workload however it seems that TRIM operations don’t occur and I'd guess that would be the same for any SSD with that workload.

That's part of what is puzzling with the performance issues I'm having. The drive acts a lot like one without TRIM, even when it's clearly working. I think there is something strange with how the drive works with static data, but I'm just grasping at straws here.

Also, is it possible 177 is just reporting maximum erase count? Could that explain the massive discrepancy between what is expected and what is actual?

[Anvil]

Ao1

That is correct, there can be no TRIM on such writes, it takes a delete to trigger TRIM and that will not occur during such I/O.
If the testfile is full-span then one would expect a TRIM to return the drive to almost fresh state when the test-file is deleted, that is not always the case though.

@minpayne

WA = NAND writes / Host writes

We can't tell for sure that we have found NAND writes for any drives with the exception of the SandForce based drives where E9 shows actual writes to NAND.
MWI Exhausted in the first chart is not an estimate, it is the actual bytes written by the Host at the time of MWI exhaustion, as in MWI = "0".
There are 2 drives (that are currently running) that shows "unexpected" results and those are the 830 and the G1 (w/o TRIM), all the other drives are within a few "cycles" of the NAND spec.

I expect we can find out what's going on with the 830 but the reporting looks to be "faulty" during this test, as it looks to require a power cycle for the SMART attributes to update.

The 830 really is strange, if we can be sure that 177 raw value = wear leveling count or "P/E count" then it would mean that it has written 13K times it's own capacity which would mean > 750TiB when Host writes shows that 200TiB has been written.


@Christopher
I was thinking the same as you , looking at the figures the 830 looks a bit like a drive w/o TRIM.
It might need GC to kick in in order for TRIM do actually do it's work and it's not allowed that pause and so eventually it slows down.

[johnw]

looking at the figures the 830 looks a bit like a drive w/o TRIM.
It might need GC to kick in in order for TRIM do actually do it's work and it's not allowed that pause and so eventually it slows down.

If that is the case (and I agree it seems like it could be), then the next obvious question is, why did Samsung change the firmware, and what exactly did they change? Since the 470 did not slow down significantly from the first hour of ASU to the last.

I think it is somewhat more likely that the slowdown is a firmware bug. If you search the web, you can find a report or two of people who have had their 830's slow down a lot, even on read speed. The read speed slow down is what makes me think it is a bug. I can see write speed slowing down due to delayed GC, but read speed really should not slow down with any reasonable design choices.

[Anvil]

Let's hope it's a bug and not a "design flaw".

I haven't noticed any slow down on my 830's but they aren't exposed to extreme conditions, I'll google the slow down.

[Christopher]

I have no doubts that the 830 would excel in a desktop workload, and you can't argue with the speed either. I'm laboring under two assumptions at this time: first, that there is something funky with the TRiM, and secondly, that 177 must be max erase count, not an average. There is just no way possible for the Samsung to have used 13K PE cycles in such a short period of time.

The Samsung is using 16PE cycles per hour if 177 is the average PE cycles. I believe it must be max erase count, or else its wrong.

[Anvil]

There is a new Magician V3.1 for the Samsung (dated 2012-01-06), it says nothing about the firmware though. Link

1. Title
Samsung SSD Magician Software Ver. 3.1

2. Applicable Model
- 470 Series: MZ-5PA064* MZ-5PA128* MZ-5PA256*
- 830 Series: MZ-7PC064* MZ-7PC128* MZ-7PC256* MZ-7PC512*

3. Improvements & Modifications
- Added CD/DVD burning tool for Firmware update & secrue erase
- Fixed minor bugs

4. Operating System
- Windows XP/Vista/7 (Recommended)

* Notice: SSD will not be recognized while in RAID array mode.

[B.A.T]

Todays update.
Kingston V+100
329.4877 TiB
1823 hours
Avg speed 25.02 MiB/s
AD still 1.
168= 1 (SATA PHY Error Count)
P/E?
MD5 OK.
Reallocated sectors : 00


Intel X25-M G1 80GB
220,3054 TiB
20132 hours
Reallocated sectors : 00
MWI=141 to 140
MD5 =OK
49.81 MiB/s on avg


m4
205.3534 TiB
752 hours
Avg speed 80.43 MiB/s.
AD gone from 240 to 236.
P/E 3610.
MD5 OK.
Reallocated sectors : 00

[Christopher]

What are the stats on the 320? Are reallocations still holding steady?

Despite the initial speed advantage, the 830 and M4 are neck and neck still. But from here on out, the 830 will most likely be surpassed by a considerable margin over the next month.

[Ao1]

Average WA
60GiB (LBA) * P/E Cycles (499) = 29,940
Actual writes = 2,498
WA = 11.98

MWI 86
POH 65

WA is down, which reflects the time I spent yesterday creating and deleting 4 MiB sequential files to test trim.