Best SSD for heavy use without TRIM?

[SgtSpike]

I run a Minecraft server off a dedicated homebuild machine. For those who do not know what Minecraft is, it is a PC video game with infinite modifiable maps made of blocks. These maps are currently stored in "chunks" of 16x16x128 blocks, which end up being individual 4KB files.

An update to the game will fix this horrible filesize inefficiency (think about what near constant read/writes of 5GB of 4KB files would do to your typical SSD), bringing the file size for each chunk up to 4MB.

While the files were stored in 4KB chunks, I attempted to stored the map files for the two servers I run on an SSD, a Vertex 1 60GB. Needless to say, I had about 6500 of the 10000 read/write cycles already used up after a period of just two months of using the drive in this manner. I figured this was because the drive had to erase an entire page to write a 4KB file once the drive was more than... 5% full? Not sure exactly. Depends on the page size, and I'm not sure what it is on the Vertex 60GB. Somewhere between 512KB and 2MB though, from what I have read.

At any rate, the new 4MB/chunk map file format should greatly extend the life of any SSD, to the point where it would be practical to purchase one for using in the server.

I run Server 2008 (non-R2), so I do not have TRIM support. I cannot afford R2 ($500+), so I would like to know what the best SSD is to use when TRIM support is not available, and when the drive is expected to have heavy read/write activity.

I was thinking the OCZ Vertex 2, as an AnandTech review shows it losing very little of its performance after being "used", but recent reviews on Newegg seem to imply that OCZ is suddenly using inferior, lower-speed NAND for those drives, which makes me nervous to rely on them. Are there any other reasonably-priced drives that are still decently fast without TRIM support? Or should I go with the Vertex 2 despite the reviews?

For an idea of the level of activity, I wrote 11TB of 4KB data to the 60GB drive I had in 2 months.

 

[LiuKangBakinPie]

I run a Minecraft server off a dedicated homebuild machine. For those who do not know what Minecraft is, it is a PC video game with infinite modifiable maps made of blocks. These maps are currently stored in "chunks" of 16x16x128 blocks, which end up being individual 4KB files.

An update to the game will fix this horrible filesize inefficiency (think about what near constant read/writes of 5GB of 4KB files would do to your typical SSD), bringing the file size for each chunk up to 4MB.

While the files were stored in 4KB chunks, I attempted to stored the map files for the two servers I run on an SSD, a Vertex 1 60GB. Needless to say, I had about 6500 of the 10000 read/write cycles already used up after a period of just two months of using the drive in this manner. I figured this was because the drive had to erase an entire page to write a 4KB file once the drive was more than... 5% full? Not sure exactly. Depends on the page size, and I'm not sure what it is on the Vertex 60GB. Somewhere between 512KB and 2MB though, from what I have read.

At any rate, the new 4MB/chunk map file format should greatly extend the life of any SSD, to the point where it would be practical to purchase one for using in the server.

I run Server 2008 (non-R2), so I do not have TRIM support. I cannot afford R2 ($500+), so I would like to know what the best SSD is to use when TRIM support is not available, and when the drive is expected to have heavy read/write activity.

I was thinking the OCZ Vertex 2, as an AnandTech review shows it losing very little of its performance after being "used", but recent reviews on Newegg seem to imply that OCZ is suddenly using inferior, lower-speed NAND for those drives, which makes me nervous to rely on them. Are there any other reasonably-priced drives that are still decently fast without TRIM support? Or should I go with the Vertex 2 despite the reviews?

For an idea of the level of activity, I wrote 11TB of 4KB data to the 60GB drive I had in 2 months.

Why not use a raid array with a pci-e raid controller. Get a couple of Samsungs hook em up. We do it in Adobe as scratch disks. Theyre performance are excellent

 

[Emulex]

i'd skip sandforce for any server application. having a drive not come up in a raid scenario can be painful. sandforce has some clever timing issues 🙂

 

[SgtSpike]

Why not use a raid array with a pci-e raid controller. Get a couple of Samsungs hook em up. We do it in Adobe as scratch disks. Theyre performance are excellent

Because a single SSD would still be faster, and I don't need a lot of space. And, I don't really like RAID0.

Even though the mapfile chunk size is changing to 4MB, it's still going to have a LOT of individual files (not all of the previous 4KB chunks will match up to the same larger-size chunks). I need the random access time to be as quick as possible to ensure the HDD doesn't bog the server down. RAID doesn't help random access times.

I am currently using a VelociRaptor drive to store one of the servers on. It is horribly slow, and has given me faith that even the worst of SSD's would be scores better than any HDD for this particular application. That might change with the new mapfile size, so I will do more testing and analysis once that change is made.

i'd skip sandforce for any server application. having a drive not come up in a raid scenario can be painful. sandforce has some clever timing issues 🙂

It won't be in RAID. Does that change your opinion?

 

[Emulex]

just pay attention for the drive dropping on cold boot or sleep. if so the firmware is not up to par. many sandforce drives are running older firmware with no updates. that is my problem. the adata s599/microcenter G2 actually works 99.9% of the time with the newer 3.4.6 firmware which is far better than my callisto deluxe

 

[SgtSpike]

Ok, will do. It won't be sleeping, and won't be doing a cold boot unless I'm installing/upgrading hardware.

Regardless, what would be your recommendation for a drive that performs well even when dirty without TRIM support? That's the heart of my question...

 

[Mark R]

The problem with 4k files is that they may not benefit much from TRIM. Many drives ignore TRIM requests smaller than a certain size limit - which can be as high as 256 k. So you TRIM is only really worthwhile where the files are larger. Additionally, TRIM only works where files are deleted (not where they are modified) - so if minecraft modifies files, rather than deleting them and creating new ones - there would be no benefit from TRIM anyway.

Your figures show a write amplification of about 35x. This is pretty normal for an OCZ Vertex. These drives use an indilinx controller which has a very bad write amplification factor, giving a very short drive life span. I've seen some indilinx drives claim WA figures of over 70x from normal desktop use - but those were the very low capacity ones with less overprovisioning. The WA performance of the indilinx controller is between 10-30x worse than competing controllers from Intel or Sandforce.

To work around the SSD wear issue, it's best to use a drive with as large a capacity as is reasonable. This spreads the wear over a larger number of cells. You should also manually overprovision some space (by secure erasing the drive and leaving part of the drive unpartitioned). By and large, once you get about 20-25% overprovisioning, there's little point in going further. Good controllers with lots of overprovisioning (10+ GB) should be able to handle lots of 4k writes without too much WA.

Another option to minimze flash writes is to use a SAS/SATA card with battery-backed write cache optimized for SSD use. These cards can hold writes in battery-backed RAM, until suitably sized batches of writes can be made. However, the price of these cards seems to make them impractical for your use.

Some people have produced special drivers that emulate the same effect in software - they hold small (<64k) writes in RAM in an attempt to find multiple writes that are adjacent, and combine them into a single write. (e.g. flashfire) The risk with delaying writes like this, however, is that you run the risk of catastrophic drive corruption if the power goes out or the OS crashes.

Realisticaly, your best option is to find a large SSD with a sophisticated controller (Sandforce or Intel) ideally with supercapacitor backup, to permit more aggressive write aggregation. Additionally, the use of 34 nm flash is preferable due to its higher endurance.

 

[jiffylube1024]

I'd go with Intel for the reliability. If you can afford an X25-E then they're among the most reliable drives on the market AFAIK, otherwise an X25-M would be a wise choice as well.

 

[SgtSpike]

jiffylube, the write speed on the Intel X25-M drive is pretty poor to start with... do you know anything about how much performance drops when it handles dirty cells?

Mark R, thanks for the input. I did not know about controllers ignoring TRIM on small file sizes, nor did I really think about the fact that TRIM would be unnecessary unless the whole cell was cleared of files (which would be fairly unlikely in my case). So it looks like my choice of OS will not matter much anyhow.

Can you explain more about write amplification? Does this mean the number of bytes of files written versus the number of bytes cycled?

Does secure-erasing need to be done to overprovision even on a brand new drive? And a Sandforce controller would potentially give me 10-30x the life out of the SSD? That would be plenty, as it would ensure the SSD lastest at least a year.

Right now, I am considering the Mushkin Callisto Deluxe. From what I understand, it has the same Sandforce controller as the Vertex 2, claims the same speeds, but still uses the 34nm NAND.

 

[Cable God]

Check out the OWC RE Pro, for the 28&#37; over-provisioning and 5yr replacement warranty, even if you "write out" the drive.

 

[Emulex]

owc is another ocz is another ... (get the picture).

 

[Cable God]

True, sandforce is sand force, it has it's short-comings like everything else. I think his requirements are easily met with the OWC RE though. I've had several OCZ failures, but none from OWC so far. I have about 50 OWC RE's deployed since about 5 months ago or so, and they have been nothing short of great. At least if he writes it out every 6 months, it will be replaced next-day every time, no questions asked. * No, I have no vested interest in OWC besides being a customer.*