Sign in to follow this  
DigitalFreak

Sandisk 32GB 2.5" SATA SSD drive...

Recommended Posts

Really nice! 7000 IOPS!!! But did anybody find anything about Write speed? Anyway for small databases a few of those (perhaps they are reliable enough to skip mirroring when only stripping 3-4 drives?) would be VERY nice :).

For home use I'd buy two if I could get 16GB with identical performance at half the price, just to play around with them a little with a database and then use them as system drives.

I guess retail price will be at least 400 though.

Share this post


Link to post
Share on other sites

IIRC the write speed was 42 or 45 MB/sec. Not bad at all! Sandisk has a PDF spec sheet with all the details. I just have one question: When can I buy one?

Share this post


Link to post
Share on other sites

Looks like a "paper launch".

Everywhere you look are copies of the Sandisk press-release and promises that the product is launched, and available to OEM's.

However I have not yet found a review here at SR, or AnandTech or any other reputable reviewer. If it's not available to them, I doubt it's going to be available to you and I anytime soon. This will be a more interesting story when it's less repetition of the press release fluff, and more meat.

Share this post


Link to post
Share on other sites

The performance of Samsung’s SSD exceeds that of a similarly sized hard drive by more than 150%. For example, it can read data at a rate of 56MB per second and achieves write speeds of 32MB per second

full note on the samsung site here:

http://www.samsung.com/eu/products/semicon...roducts/ssd.asp

I want 100mb/s read speed :D If i'm going to buy one, don't really want to go slower than my raptor, even if the seek is nearly non-existant.

Nox

Share this post


Link to post
Share on other sites

I'm hoping for 16GB versions at resonable prices, 2 of them will get you past 100MB/s read. If the reliability is as good as they claim raid 0 with a few drives is going to be an ok risk to take.

Share this post


Link to post
Share on other sites

ah, i should of said 'on each' :D I'm happy to wait till xmas, think i'll be upgrading my gfx card first anyway.

Nox

Share this post


Link to post
Share on other sites

You can get fast SSDs already -

Lexar and Sandisk have already launched CF cards that support UDMA. Sandisk's is shipping; Lexar's is a paper launch.

The speeds are very good. Lexar is claiming sustained writes of 45MB/sec and a launch capacity of 8GB. The Sandisk Extreme IV 8GB CF card claims 40MB/sec read/write. I found it in stock for $229 and there is even a $20 rebate being offered, for $209 delivered. It's nowhere near the $/GB of the SSDs they just announced, but you can actually get your hands on one.

The beauty of the UDMA CF cards is that you can stick them in dirt cheap CF/IDE adapters and use them to their full potential. I've even seen a CF/SATA adapter for sale.

I wish I could replace the agonizing 1.8" HDD in my Sony laptop with a CF card!

Share this post


Link to post
Share on other sites

I'm thinking that these drives (or similar drives) might be useful to form some intermediate storage level between primary and secondary storage (although these drive reasonably should be considered secondary storage).

Since these flash based storage allows very high amounts of IOPS but are a bit to expensive for entire databases (at least the ones I work with). An intermediate solution could be to have a raid 0 array with a few of these drives and only store data that can be easily recreated if necessary (like indexes).

It would offload the main storage from dealing with all index related IO which is often many small reads, perhaps the performance gained would be great for a relative modest cost.

Anybody else been considering this?

Share this post


Link to post
Share on other sites

Actually, index I/O (at least with SQL server) is mostly sequential 1MB..so flash would actually be a hinderence here. I can dream about a flash based TPC-C run though. If we're indeed looking at 7000 8k IOPS, that means each flash drive could essentially replace 35 15k SAS drives..which is INSANE.....:-)

I'm thinking that these drives (or similar drives) might be useful to form some intermediate storage level between primary and secondary storage (although these drive reasonably should be considered secondary storage).

Since these flash based storage allows very high amounts of IOPS but are a bit to expensive for entire databases (at least the ones I work with). An intermediate solution could be to have a raid 0 array with a few of these drives and only store data that can be easily recreated if necessary (like indexes).

It would offload the main storage from dealing with all index related IO which is often many small reads, perhaps the performance gained would be great for a relative modest cost.

Anybody else been considering this?

Share this post


Link to post
Share on other sites
Actually, index I/O (at least with SQL server) is mostly sequential 1MB..so flash would actually be a hinderence here. I can dream about a flash based TPC-C run though. If we're indeed looking at 7000 8k IOPS, that means each flash drive could essentially replace 35 15k SAS drives..which is INSANE.....:-)

I'm not sure but the 7000IOPS where specified for 512B reads i think so it won't be that fast for the standard 8k blocks, with any luck it would do 7000/16=400IOPS which is still impressive, compared to a sata drive it is magnificent!

Hmm... just contemplating what you're saying above... for index scan operations I think you are right but for index seeks, wouldn't IO likely be mainly small block requests? I think normally to pin point a particular tuple it takes 2-4 block accesses in the index and then a trip to the table itself to use the primary key to find the right tuple.

Of course the entire database on drives like this would be crazy, but I think I first have to display a performance increase for the database we are currently working with (over a TB) to make the boss fork up some money. I'm thinking that using these drive to store table data is not very attractive since that would cut the number of effectively used drives in half with raid1(0) (and putting production data on non redundant storage is out of the question, even putting indexes on non redundant storage is pushing it but could be done during an evaluation).

On the other hand a considerable amount of small tables could be fitted onto a pair of those drives and any offload on the main array should give a noticable performance increase. Considering STR large tables really needs to stay on large disk arrays, or reports and backups will take all night :(.

Another thought... for queries where the expected amount of tuples selected reach some threshold table scans are usually performed instead of using the index to pick each tuple one at a time (because of the slow seeks). With flash based storage the point where table scans are preferred should be pushed a little further I wonder if this possible to tune?

Share this post


Link to post
Share on other sites

BTW... are current controllers ready for these drives? It seems likely that getting the most out of say 8 drive array you might have to buy a 16 or 24 port controller just to get a controller with a processor capable of pushing the drives.

(I must admitt that I would be pretty interested in seeing a small database server with 24 of these babies, over 300GB of near instant access data, it would take an extra cpu-core just to keep the drives busy :))

Share this post


Link to post
Share on other sites
Looks like a "paper launch".

Everywhere you look are copies of the Sandisk press-release and promises that the product is launched, and available to OEM's.

However I have not yet found a review here at SR, or AnandTech or any other reputable reviewer. If it's not available to them, I doubt it's going to be available to you and I anytime soon. This will be a more interesting story when it's less repetition of the press release fluff, and more meat.

With all due respect to SR, they tend to be a fairly traditional as far as the products they choose to review. There' have been some 'outside of the box' type storage products that have come out over the years and I always hoped SR would weigh in with their analysis, but they seem to just stick with traditional mechanical hard drives for the most part.

I'm looking forward to real availability too. It's a breakthrough price for what it offers, but still not an impulse buy, so hopefully we get some good reviews. The main attractions for me are light weight, cool running, and zero noise. If I can get all that and it matches current HD speeds, I'm sold.

As far as the availability of CF cards and cheap IDE/SATA adapters, I looked into it and heard the current bridge chips were designed for lower speed CF cards and end up being a severe bottleneck when trying to use them as a standard hard drive. Just what I heard, mind you, which is why it would be nice if SR would do their typical in depth analysis on these fringe types of storage technology as well.

Share this post


Link to post
Share on other sites
As far as the availability of CF cards and cheap IDE/SATA adapters, I looked into it and heard the current bridge chips were designed for lower speed CF cards and end up being a severe bottleneck when trying to use them as a standard hard drive. Just what I heard, mind you, which is why it would be nice if SR would do their typical in depth analysis on these fringe types of storage technology as well.

CF to IDE adapters are completely passive devices. CF cards are native IDE in the first place. (They also support other modes, but that's of no consequence to us.) The adapters just give you a way to plug them in. There are no bridge chips for CF-IDE. CF-SATA would use the same bridge chip (i.e., Marvell) that early-generation SATA drives used. There is little to no performance loss with such a bridge chip.

The catch is that nearly all existing CF cards only support PIO transfers, and that is a huge bottleneck. The other problem is that Flash has always been a slow technology, but it's been gaining speed quickly. That's why the new UDMA cards are a big deal. They eliminate the PIO bottleneck, use multi-channel flash controllers, write buffers, etc., and should outperform magnetic discs on just about any real workload (not on sequential transfer rate tests).

Share this post


Link to post
Share on other sites

Pretty sure that any CF card rated above 16MB/sec has to be using DMA, since that's the best possible speed for PIO.

Interestingly enough, it appears that CF to Cardbus adapters are not purely passive; only Delkin seems to make one that supports DMA and it's more expensive than most other CF to PCMCIA adapters.

Share this post


Link to post
Share on other sites
Pretty sure that any CF card rated above 16MB/sec has to be using DMA, since that's the best possible speed for PIO.

Oops, that is correct. A handful of CF cards support DMA modes 1, 2 or 4. The new cards are the first that support UDMA. Sandisk Extreme IV uses UDMA 4.

Unfortunately the cheaper 120X CF cards suck at non-sequential small writes.

Interestingly enough, it appears that CF to Cardbus adapters are not purely passive; only Delkin seems to make one that supports DMA and it's more expensive than most other CF to PCMCIA adapters.

Cardbus is a derivative of PCI, so Cardbus to CF adapters would need to have a PCI/IDE controller chip. That Delkin card is pretty cool. I forget who makes it. Delkin just slaps their name on it.

Most of the cheaper adapters are PCMCIA, not Cardbus. I'm not sure if they are completely passive but they are definitely much simpler.

Share this post


Link to post
Share on other sites

Hi!

I just got my delkin reader and without being too enthusiastic it has changed my view on cf cards. Just to test it I also got a Sandisk ultra 3 card that promises 20MB/s read and write I think (not sure about the numbers).

Anyway I plugged in the reader and it was a hazzle to get it going due to drivers but I finally found a guide on the internet and got it working...

Filled the card with files and read them back and it manages over 10MB/s (read and write) (compared to the 3-4MB/s with other various readers). Also compared to my previous cardbus adapter cpu-load is very low (instead pegged att 100% for one of the cores)

Just to test it I created a 3.5GB file (4GB - some files I already stuffed it with :)). Wrote a simple program to run 5000 reads with some data sizes. Ran the test on my laptop 7200rpm drive and then on my CF card. (Note it is not a very good test since 3.5GB is only a small portion of the hard drive in my laptop but I had to run it on something to compare with.

The numbers...

HDD 64KB: 10.93ms

HDD 16KB: 10.93

HDD 8KB: 7.78ms

HDD 1KB: 7.86ms

CF 64KB: 4.83ms

CF 16KB: 1.84ms

CF 8KB: 1.35ms

CF 1KB: 1.06ms

As expected the HDD is not very sensetive to the data size, main factor is to get started reading. The CF card is very limited by its low transfer speed so data size affects performance much. It still manages some 700 reads per second using 8KB blocks which is far above what can be expected by any enterprise drives today.

This is a simple CF-card in a cardreader (although an above average card with an excellent reader). Still I think that as soon as SSD drives arrive and if they can at least deliver the IO and transfer numbers promised so far, they will forever change the storage systems at least for OLTP databases. However due to limited transfer speeds I think that some tuning is required to make the database optimiser avoid table scans.

Share this post


Link to post
Share on other sites
CF to IDE adapters are completely passive devices. CF cards are native IDE in the first place. (They also support other modes, but that's of no consequence to us.) The adapters just give you a way to plug them in. There are no bridge chips for CF-IDE. CF-SATA would use the same bridge chip (i.e., Marvell) that early-generation SATA drives used. There is little to no performance loss with such a bridge chip.

The catch is that nearly all existing CF cards only support PIO transfers, and that is a huge bottleneck. The other problem is that Flash has always been a slow technology, but it's been gaining speed quickly. That's why the new UDMA cards are a big deal. They eliminate the PIO bottleneck, use multi-channel flash controllers, write buffers, etc., and should outperform magnetic discs on just about any real workload (not on sequential transfer rate tests).

I hate to be the one to bump up a month old topic, but I'm one such user who is currently using an extreme 4 card as my primary boot device. It's small, only 4 GB, but being as I run linux it's more than capable for the job.

I've posted a ton of data over here:

http://www.phoronix.com/forums/showthread.php?t=3424

Screenshots here:

http://www.angelfire.com/rpg2/tweakit/index.html

I plan on creating another discussion when I upgrade to an 8GB card. I'll probably go with a lexar 300x card instead. I haven't been able to get full performance out of this, I wish I could find someone who knew more about this and could help out. I'm hoping that going with the lexar professional udma card might change the results I'm seeing; we'll see. I think it's primarily a driver issue but I ultimately don't know.

But it is doable. And it is faster than my previous hard drive, even running at a crippled speed. I'll never go back. Hard drives are strictly a place for various data to me now.

Edited by halfmanhalfamazing

Share this post


Link to post
Share on other sites

A few monthes ago, I got a HAMA SSD in germany (which AFAIK is a OEM sandisk SSD). I wanted to check performance for Oracle database. In a typical configuration Oracle uses 8k IO's.

Read performance was superb of course; i got roundabout 3000 IOPS with 8k blocks and 55mb/s at large block reads.

But write performance was just a mess. At 1-2MB block writes, i got 15MB/s . Writing 8kb blocks, which Oracle uses for tablespace files and redo log writes, i was not able to get more than 55-60 IOPS. So, i decided to return this thing and stick to my Areca SATA raid (which are not even Raptors!)

Now, I'm decided to go for the ARC-1680, 512 or 1024MB Cache module combined with some 15k.1 or 10k.2 RPM Savvio's......

Edited by sasfan

Share this post


Link to post
Share on other sites

After glancing over this thread and hitting the Sandisk links I started going to other common Memory manufacturers. Realized a lot of players are getting in this game. PNY, PQI, A-data, Team Group, among others. I'm very glad more people are jumping in. That means the demand for cheap NAND flash is rising.

On the down side the supplying end seems to still be very quiet and has been all year. Samsung announced a 32Gb chip last year. K9WBG08U1M's are in the wild but there has been no talk of a 64Gb chip. Hynix and Micron don't even have 32Gb chips yet...

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now
Sign in to follow this