Overclock.net › Forums › Components › Hard Drives & Storage › SSD › Samsung 840 EVO read speed drops on old-written data in the drive
New Posts  All Forums:Forum Nav:

Samsung 840 EVO read speed drops on old-written data in the drive - Page 309

post #3081 of 3279
Quote:
Originally Posted by Frag1le View Post

A few people asked me if i could do a follow up:



"The database on this array"...

Isn't this database also being written to? Wouldn't that mean the drive is slowing more as a result of fragmentation as opposed to the (fixed) stale data issue?
post #3082 of 3279
Quote:
Originally Posted by malventano View Post

Isn't this database also being written to? Wouldn't that mean the drive is slowing more as a result of fragmentation as opposed to the (fixed) stale data issue?

No first off: SSDs do not suffer from fragmentation. The database itself can suffer from fragmentation but you won't notice that with a benchmark that doesn't access the database. The benchmark just reads the big blob sequential.
SSDs aren't disks, with SSDs data is written to the memory and the controller keeps track of which bits in the memory are used to store a file.

Also as i've written before, this database is read intensive and not so much data is written to it. If there was data written to this database continuously then the data on the array would've been newer and therefore i probably wouldn't be able to show you almost any degradation.
post #3083 of 3279
Quote:
Originally Posted by Frag1le View Post

No first off: SSDs do not suffer from fragmentation. The database itself can suffer from fragmentation but you won't notice that with a benchmark that doesn't access the database. The benchmark just reads the big blob sequential.
SSDs aren't disks, with SSDs data is written to the memory and the controller keeps track of which bits in the memory are used to store a file.

Also as i've written before, this database is read intensive and not so much data is written to it. If there was data written to this database continuously then the data on the array would've been newer and therefore i probably wouldn't be able to show you almost any degradation.

First off, you are incorrect. SSDs absolutely do fragment at the flash translation layer. You are using a consumer class SSD for a database workload (not its intended purpose). Firmware on consumer SSDs is not optimized to handle small random writes across a large span (half the drive in your case). Even if the writes are rare, just writing randomly *once* across half the span of a consumer SSD will slow it considerably (including read performance). Here is some additional information on this type of behavior from a 6-year old article:

http://www.pcper.com/reviews/Storage/Long-term-performance-analysis-Intel-Mainstream-SSDs

To state what you are likely seeing more clearly, due to the FTL and random writing over time, your HDTune pass that is reading LBA's sequentially, is actually reading from the flash randomly.

Source: I wrote the article above, and I have been reviewing SSDs for nearly a decade.

Allyn Malventano
Storage Editor, PC Perspective
post #3084 of 3279
Quote:
Originally Posted by malventano View Post

First off, you are incorrect. SSDs absolutely do fragment at the flash translation layer. You are using a consumer class SSD for a database workload (not its intended purpose). Firmware on consumer SSDs is not optimized to handle small random writes across a large span (half the drive in your case). Even if the writes are rare, just writing randomly *once* across half the span of a consumer SSD will slow it considerably (including read performance). Here is some additional information on this type of behavior from a 6-year old article:

http://www.pcper.com/reviews/Storage/Long-term-performance-analysis-Intel-Mainstream-SSDs

To state what you are likely seeing more clearly, due to the FTL and random writing over time, your HDTune pass that is reading LBA's sequentially, is actually reading from the flash randomly.

Source: I wrote the article above, and I have been reviewing SSDs for nearly a decade.

Allyn Malventano
Storage Editor, PC Perspective

I never wrote that SSDs do not fragment, i wrote they do not suffer from fragmentation( in the way HDDs suffer from fragmentation).
All SSDs i use for databases are Enterprise grade, just like the 845 DC EVO's are (the DC stands for Data Center). None of the SSD arrays i manage suffer from read degradation as much as these 845 DC EVO's with sequential read benchmarks with a fixed block size.
This again isn't caused by fragmentation but caused by the voltage drops in the TLC.
post #3085 of 3279
Quote:
Originally Posted by Frag1le View Post

I never wrote that SSDs do not fragment, i wrote they do not suffer from fragmentation( in the way HDDs suffer from fragmentation).
All SSDs i use for databases are Enterprise grade, just like the 845 DC EVO's are (the DC stands for Data Center). None of the SSD arrays i manage suffer from read degradation as much as these 845 DC EVO's with sequential read benchmarks with a fixed block size.
This again isn't caused by fragmentation but caused by the voltage drops in the TLC.

Understood (now) that you are running on DC series SSDs. I was mistaken in my pror posts as I assumed you were using the consumer 840 EVO SSDs as that was what this thread was titled. Since you are on SSDs that may be susceptible to the slow down issue, I'm now inclined to say that what you are seeing is the slow down, and that your data set is likely very rarely written to (not enough to fragment and impact subsequent read performance). The fact that written data refreshes the speed combined with your results backs that up further.

Now circling back, *all* SSDs 'suffer from fragmentation', especially when writing data randomly to a file and then attempting to read that file back sequentially (which is actually reading from the flash randomly). I am currently testing 845DC SSDs here (among other enterprise units), and just like any other SSD, they do slow as they are randomly written. Enterprise series SSD / firmware handles this better than consumer, but there is indeed still a slow down.

*edit* - to double clarify - when read speed is lower than it should be (reading sequential LBA's but reading from the flash randomly and therefore more slowly), that is 'suffering from fragmentation'.
Edited by malventano - 9/18/15 at 2:33pm
post #3086 of 3279
Ok, I'm still having issues with my 840 EVO on the EXT0DB6Q firmware.
Background:
HP workstation with Xeon W3570, Dual GPUs/various PCI-e devices, so the case runs pretty hot (important later)
Drive is on a PCI-e controller card, and performance appears to be capped at 200MB/sec
This workstation has daily use of general read/writes and old data as well
AV was turned off for the tests
Note: the drive was freshly cloned in mid-July, but I only have data starting in September.

When I started:

Warning: Spoiler! (Click to show)

Current-day:


Results:
* 197MB/sec vs 164 = ~17% decrease over two weeks!!
* More variance with current test results
* Case runs really, really hot--the case only has a single 92mm exhaust, plus the PSU fan. No intake, and the ambient goes up to 82F at night. CPU is a pretty hot chip. I read a few pages back that higher drive temps induce a more rapid degradation of old data, and this seems to be the case.

Also, I should figure out what is bottlenecking my drive speed. I'm guessing it must be to shared PCI-E lanes or something.
I'm not sure how to proceed. All signs point to drive degradation, but I'm not sure who is left to represent me, since I bought it back in Feb 2014 from Newegg.
Edited by Breex243 - 9/30/15 at 3:11pm
post #3087 of 3279
Quote:
Originally Posted by Breex243 View Post

Ok, I'm still having issues with my 840 EVO on the EXT0DB6Q firmware.
Background:
HP workstation with Xeon W3570, Dual GPUs/various PCI-e devices, so the case runs pretty hot (important later)
Drive is on a PCI-e controller card, and performance appears to be capped at 200MB/sec
This workstation has daily use of general read/writes and old data as well
AV was turned off for the tests

When I started:

Warning: Spoiler! (Click to show)

Current-day:


Results:
* 197MB/sec vs 164 = ~17% decrease over two weeks!!
* More variance with current test results
* Case runs really, really hot--the case only has a single 92mm exhaust, plus the PSU fan. No intake, and the ambient goes up to 82F at night. CPU is a pretty hot chip. I read a few pages back that higher drive temps induce a more rapid degradation of old data, and this seems to be the case.

Also, I should figure out what is bottlenecking my drive speed. I'm guessing it must be to shared PCI-E lanes or something.
I'm not sure how to proceed. All signs point to drive degradation, but I'm not sure who is left to represent me, since I bought it back in Feb 2014 from Newegg.

Well a few things could be happening.

1. if your capped at 200 MB/S, you're probably on a SATA level 2 controller

2. Evo's have something called Dynamic Thermal Guard, so if the SSD passes a certain threshold, the controller throttles the speed of the drive to prevent failure.

3. you might be on a secondary controller or a pci-e adapter if its using PCI-E lanes, i would recommend to switching to the Primary SATA controller on the board.
My Stuff
(13 items)
 
  
CPUMotherboardGraphicsGraphics
Intel Core i7-6700K ASUS Z170-A  EVGA GeForce GTX 980 K|NGP|N EVGA GeForce GTX 980 K|NGP|N 
RAMHard DriveCoolingOS
G.SKILL Ripjaws 4 Series 16GB (4 x 4GB) Samsung 850 Pro 256GB Corsair Hydro Series H100i GTX Windows 10 Pro 
MonitorKeyboardPowerCase
AOC G2460PG 24" G-SYNC 1ms (GTG) 144hz  Logitech Wireless Illuminated Keyboard K800 Antec HCP-1000 Platinum thermaltake chaser series chaser mk-i 
Mouse
Logitech G700S 
  hide details  
Reply
My Stuff
(13 items)
 
  
CPUMotherboardGraphicsGraphics
Intel Core i7-6700K ASUS Z170-A  EVGA GeForce GTX 980 K|NGP|N EVGA GeForce GTX 980 K|NGP|N 
RAMHard DriveCoolingOS
G.SKILL Ripjaws 4 Series 16GB (4 x 4GB) Samsung 850 Pro 256GB Corsair Hydro Series H100i GTX Windows 10 Pro 
MonitorKeyboardPowerCase
AOC G2460PG 24" G-SYNC 1ms (GTG) 144hz  Logitech Wireless Illuminated Keyboard K800 Antec HCP-1000 Platinum thermaltake chaser series chaser mk-i 
Mouse
Logitech G700S 
  hide details  
Reply
post #3088 of 3279
Quote:
Originally Posted by Palorim12 View Post

Well a few things could be happening.

1. if your capped at 200 MB/S, you're probably on a SATA level 2 controller

2. Evo's have something called Dynamic Thermal Guard, so if the SSD passes a certain threshold, the controller throttles the speed of the drive to prevent failure.

3. you might be on a secondary controller or a pci-e adapter if its using PCI-E lanes, i would recommend to switching to the Primary SATA controller on the board.

Actual performance speed doesn't really matter for the purposes of this, as I was looking for relative/delta performance.
I know my speed is bottlenecked because I don't have an intel SATA III controller on this board/CPU.
(The long of it was in the middle of this study, I wanted to use a PCI-E card with higher max speed (~400MB/sec on this Asmedia chipset), but somehow bottlenecked my speeds even more down to 200MB/sec. Like I said, I believe this is due to limited PCI-E lanes, since I was getting a botlteneck of ~275MB/sec in July on the onboard SATA II connection.)

The detail is that it was running for 2 months at a bottleneck of 275MB/sec, then switched to a card that was capping at 200MB/sec. I said **** it, let's see how this goes, as the initial SSDReadSpeedTester showed fairly consistent read speeds (all at the bottleneck), so the "4 weeks" is really the last 4 weeks of a 12-week run.

Dynamic thermal guard isn't kicking in--the drive currently idles at 33C and doesn't get any higher during the first two minutes of the read test. I can check in how high idle temp gets when the ambient increases by 10F at night.
post #3089 of 3279
Quote:
Originally Posted by Breex243 View Post

Actual performance speed doesn't really matter for the purposes of this, as I was looking for relative/delta performance.
I know my speed is bottlenecked because I don't have an intel SATA III controller on this board/CPU.
(The long of it was in the middle of this study, I wanted to use a PCI-E card with higher max speed (~400MB/sec on this Asmedia chipset), but somehow bottlenecked my speeds even more down to 200MB/sec. Like I said, I believe this is due to limited PCI-E lanes, since I was getting a botlteneck of ~275MB/sec in July on the onboard SATA II connection.)

The detail is that it was running for 2 months at a bottleneck of 275MB/sec, then switched to a card that was capping at 200MB/sec. I said **** it, let's see how this goes, as the initial SSDReadSpeedTester showed fairly consistent read speeds (all at the bottleneck), so the "4 weeks" is really the last 4 weeks of a 12-week run.

Dynamic thermal guard isn't kicking in--the drive currently idles at 33C and doesn't get any higher during the first two minutes of the read test. I can check in how high idle temp gets when the ambient increases by 10F at night.

Well you said the case gets very hot, so that's what i based that on. Maybe its the connection you are using is causing an issue? I have my EVO in my laptop and its still running strong, my readspeedtester, last i checked was avg of 510MB/s and its been running since the FW update came out.
My Stuff
(13 items)
 
  
CPUMotherboardGraphicsGraphics
Intel Core i7-6700K ASUS Z170-A  EVGA GeForce GTX 980 K|NGP|N EVGA GeForce GTX 980 K|NGP|N 
RAMHard DriveCoolingOS
G.SKILL Ripjaws 4 Series 16GB (4 x 4GB) Samsung 850 Pro 256GB Corsair Hydro Series H100i GTX Windows 10 Pro 
MonitorKeyboardPowerCase
AOC G2460PG 24" G-SYNC 1ms (GTG) 144hz  Logitech Wireless Illuminated Keyboard K800 Antec HCP-1000 Platinum thermaltake chaser series chaser mk-i 
Mouse
Logitech G700S 
  hide details  
Reply
My Stuff
(13 items)
 
  
CPUMotherboardGraphicsGraphics
Intel Core i7-6700K ASUS Z170-A  EVGA GeForce GTX 980 K|NGP|N EVGA GeForce GTX 980 K|NGP|N 
RAMHard DriveCoolingOS
G.SKILL Ripjaws 4 Series 16GB (4 x 4GB) Samsung 850 Pro 256GB Corsair Hydro Series H100i GTX Windows 10 Pro 
MonitorKeyboardPowerCase
AOC G2460PG 24" G-SYNC 1ms (GTG) 144hz  Logitech Wireless Illuminated Keyboard K800 Antec HCP-1000 Platinum thermaltake chaser series chaser mk-i 
Mouse
Logitech G700S 
  hide details  
Reply
post #3090 of 3279
Well, my 840 EVO does not regain its "factory speed" even after secure erase. Looks like once its been starting to degrade it all downhill for TLC NAND. I have secure erased it twice with the usb created from Magician 4.7 and still the same result. Does not pass over 406 read 396 write on its own benchmark. In the beginning it was something like 530/500. Plus something is really wrong as if I start the "Performance Optimization" it goes very fast till 50% then it takes about 30-40 minutes to complete until 100%.
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: SSD
Overclock.net › Forums › Components › Hard Drives & Storage › SSD › Samsung 840 EVO read speed drops on old-written data in the drive