Saturday, November 13


Burn In... Burn Out

If you run an 80GB Intel X25-M at its rated 4KB random-write performance, it will exceed its rated 4KB random-write endurance after just 3 days.

For a 32GB Intel X25-E, it would take about 28 months.

This is why the E-series runs five times the cost per gigabyte of the M-series.

What the real-world lifespan is for either drive range I have no idea, but I aim to find out.

Update: Based on the actual activity patterns on our server, the 80GB M-series would last 2½ years, and the 32-GB E-series (which is what we are using, 3 drives in a RAID-5 array) would last about three centuries.

Posted by: Pixy Misa at 05:49 AM | Comments (11) | Add Comment | Trackbacks (Suck)
Post contains 112 words, total size 1 kb.

1 I'm beginning to wonder if this will be a problem for the SSD in the Slate. My Win7 machine accesses the HD about once a second when idle. I assume it's writing something out, presumably to the  same location every time. If the Slate's SSD is good for a million writes, it would die in 12 days.

Posted by: Steven Den Beste at Tuesday, November 16 2010 05:38 AM (+rSRq)

2 I expect Microsoft and HP to use a custom spin of Win7 that minimizes pointless writes. If you look at some of the netbook and SSD tuning pages, there are a lot of things that can be disabled in Windows to make it run better in this sort of environment.


Posted by: J Greely at Tuesday, November 16 2010 06:19 AM (fpXGN)

3 Like the indexing service, I hope.

Posted by: Steven Den Beste at Tuesday, November 16 2010 09:11 AM (+rSRq)

4 I'm pretty sure it's a million per block, not total. Unfortunately, as I understand, something like TRIM is required for the firmware to level writes properly, otherwise it cycles through same 16 blocks or so.

Posted by: Pete Zaitcev at Tuesday, November 16 2010 12:03 PM (9KseV)

5 For MLC flash, it's 10,000 cycles per block. If both the flash controller and the drivers are pessimal, you're good for 20 million writes per GB of flash.

Posted by: Pixy Misa at Tuesday, November 16 2010 02:58 PM (PiXy!)

6 Oh, and just to clarify, the workload that would kill that 80GB drive in 3 days is 6600 random 4KB writes per second.

Posted by: Pixy Misa at Tuesday, November 16 2010 03:02 PM (PiXy!)


What I was assuming was that if the same location was written a million times, the chips would be dead. Sounds like that's a huge overestimate. If it's only 10,000 the unit wouldn't last three hours.

Presumably this is something they'd have noticed during design and testing, so presumably they've got an answer -- most likely, as mentioned above, tuning Windows so it doesn't do that kind of thing.

Posted by: Steven Den Beste at Tuesday, November 16 2010 04:25 PM (+rSRq)

8 Another possibility is a robust "bad block" algorithm. That's easy to do with a disk. When blocks go bad, you put them into the bad-block list, and the drive keeps going with reduced capacity.

Posted by: Steven Den Beste at Tuesday, November 16 2010 04:26 PM (+rSRq)

9 Yep.  In fact, the drive will automatically move data around to balance out writes, and comes with reserve capacity to replace bad blocks.  A good drive might actually be 20% larger than the stated capacity.

Since there's almost no seek time on an SSD (there's a small delay to switch pages) for most uses you hardly notice this.  A badly fragmented drive might only deliver 100MB/s on sequential reads instead of 200MB/s, and that's where TRIM comes in.  Regular software defragmentation just gets remapped by the drive anyway and achieves nothing, so they had to add a drive command to do it directly.

Posted by: Pixy Misa at Wednesday, November 17 2010 01:48 AM (PiXy!)

10 The FusionIO ioDrives we use in some of our servers at my day job are basically dumb cards full of flash chips.  They don't have intelligent controllers because the CPU in a server is much faster than anything you can fit on an expansion card, so it all happens in software.  Which means you can see what it's doing. smile

Which gave me a really lousy time in December 2008 when I was trying to do a database reload and ran out of free blocks for the wear-leveling algorithm and the drive's performance plunged by 95% and I thought it was broken and had to get a replacement part shipped to Seattle on Christmas day at midnight in a blizzard.

Well, approximately.

Posted by: Pixy Misa at Wednesday, November 17 2010 01:52 AM (PiXy!)

11 Yet despite its tiny capacity the use of the enormously expensive SLC NAND flash memory means the Intel X25-E retails for a wallet crushing £505!

Posted by: devizakereskedés at Saturday, December 04 2010 02:36 AM (EIzjX)

Hide Comments | Add Comment

Comments are disabled. Post is locked.
50kb generated in CPU 0.1105, elapsed 9.6268 seconds.
56 queries taking 9.5953 seconds, 348 records returned.
Powered by Minx 1.1.6c-pink.