Slashdot Mirror
Intel Takes SATA Performance Crown With X25-E SSD
theraindog writes "We've already seen Intel's first X25-M solid-state drive blow the doors off the competition, and now there's a new X25-E Extreme model that's even faster. This latest drive reads at 250MB/s, writes at 170MB/s, and offers ten times the lifespan of its predecessor, all while retaining Intel's wicked-fast storage controller and crafty Native Command Queuing support. The Extreme isn't cheap, of course, but The Tech Report's in-depth review of the drive suggests that if you consider its cost in terms of performance, the X25-E actually represents good value for demanding multi-user environments."
This just screams dedicated database storage.
http://blindscribblings.com - Tasty pop-culture in conceptual fashion.
Is it just me or have we gone full-frontal-funnyfarm with the analogies and adjectives here?
meep
Don't you mean the door population?
Software piracy is victimless theft.
Nevermind. "Even with 100GB of write-erase per day, it'll take more than 72 years to burn through the drive." I should RTFA. But still, much room for improvement.
I have left slashdot and am now on Soylent News. FUCK YOU DICE.
It's pretty much apples and oranges. Even with batteries (which I wouldn't trust) RAM has different characteristics in power consumption, heat output, storage density etc. By the time you address these challenges you'd have... an SSD.
Plus the SSDs get their long life from having more raw storage than advertised, and dynamically shutting down dead areas and bringing in reserve areas as it ages. Your sums would have to take into account the cost of this "hidden" storage.
As an aside the best use for these things is hands down as extended cache in a storage array. One or two SSD alongside a few terrabytes of "normal" disk managed an intelligent filesystem or storage firmware can speed the whole beast up by phenomenal amounts depending on the data usage patterns. Yet the total cost of the whole storage appliance is really not much changed in relative terms. Some of the new Sun boxes are designed to work with SSDs like this, and probably from other big storage vendors as well.
What happens when the read-write cycles on this run out?
I would have liked to have seen them test this drive in a much more powerful system. I mean, a P4 with 1GB RAM, and a fairly dated chipset (955x) as the SATA controller? No one is going to put a drive like this in a system that old. I'd guess that we might see different results on a more powerful system. At some point in those tests, other components of this fairly slow (by today's standards) machine. Throw some serious power behind it, and you can be sure that you're not bottlenecked, and the full power of the drive shows. Can't say for sure if this is actually the case, as I don't have a drive to test, but it's a definite possibility. Hopefully someone else does a similar review with a more powerful testbed.
Try using a RAM disk.
Products that use RAM as the storage media have been around for years. They're exactly what you're describing: a few standard DDR DIMMS and a battery on a PCI card, usually. However, no one in an enterprise environment would actually trust data to such a device, and they never really took off. Home users don't generally have the power and data backup capacity to safely use such a device (and not even the most hardened masochist wants to reinstall or restore everything whenever a breaker goes), and enterprise users can't tolerate the risk level. Sure, you can have backup power, but the risk of losing data and downtime restoring it just isn't tolerable in most enterprise environments.
Except those same drives don't exactly compare due to a poor implementation of the hardware or the write/cleaning algorithm in the JMicron controller many (all?) of those are using. The capacity and price are tempting, but the write latency especially during random accesses is beyond awful. Unless of course they were able to update the firmware on those chips to address the issue since this article was published: http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3403&p=7
Even if it's solid state, there's still a physical layer, and still a logical-to-physical abstraction that an IDE disk must perform. (Slashdot pedantics will please note that here an IDE disk means a disk with an integrated electronic controller, not just a drive with an ATA interface. If you've never had to know the true physical geometry -- the number of cylinders, heads, and sectors in a disk (CHS)-- to tell your PC's BIOS or OS, you've never used a non-IDE disk. Most BIOS systems were faking CHS numbers by the time EIDE hit in 1994 which eliminated CHS in favor of LBA.)
Flash drives use NAND flash memory, which uses pages of up to about 4 KB. For the most part, you can only access a single page at a time. Additionally, sequential access within a page is almost always faster than random access. Giving the disk's integrated controller a list of values means that it can examine the queue intelligently and can perform paging operations more intelligently.
The road to tyranny has always been paved with claims of necessity.
From the article: "The storage controller is an Intel design that's particularly crafty, supporting not only SMART monitoring, but also Native Command Queuing (NCQ). NCQ was originally designed to compensate for the rotational latency inherent to mechanical hard drives, but here it's being used in reverse, because Intel says its SSDs are so fast that they actually encounter latency in the host system. It takes a little time (time is of course relative when you're talking about an SSD whose access latency is measured in microseconds) between when a system completes a request and the next one is issued. NCQ is used to queue up to 32 requests to keep the X25-E busy during any downtime between requests."
Anyone who tells you SSDs are a replacement for disks is at best talking about some niche workloads and at worst trying to sell you a line of the old BS. SSDs render 15k rpm disks obsolete all right, but not in the way you're suggesting.
To get the same capacity as you'd get out of those hot, expensive disks - which is not a lot given what you're paying for them - you'd need to spend much more, and you'll likely find that performance levels off quickly when you saturate out your HBAs and/or CPUs and/or memory bus and/or front-end connectivity. Much better to combine a few of these with slow, cool, cheap disks to maximise both performance and capacity at a lower price than the 15k disks.
Let's take an example.
Suppose you have 14 146GB 15k rpm disks. They cost you $2000 apiece, or $28000. Each one gives you 300 IOPS, for a total of 4200 (we'll ignore the costs and inefficiencies of the hardware RAID dinosaur you're probably using these with; if we didn't, you might start to feel stupid about it). So you spent about $6.67/IOPS or $14/GB, plus the power and cooling to keep those disks spinning. Not cheap. Not particularly fast. Not really great in any way.
Suppose instead that you want to replace them with these 80GB SSDs. You'll probably pay your vendor around $1400 for them (figure 60% margin like they're getting on those FC drives you've been buying from them). Now you need 26 of them to get the same capacity, costing you $36400. But you get about 12000 read IOPS each (write latency suspiciously omitted from this fluff piece, but we'll dubiously assume it's similar - it almost certainly isn't anywhere close) for a total of 312000. Too bad your HBA can do only about 140000, so you'll max out there on random reads. And if we're talking about block sizes larger than 512 bytes, latency will be higher. So you've spent $0.26/IOPS, which is great, and you've saved money on operating costs as well. But you actually spent a lot more in total - $18/GB - and woe unto you if you need more capacity; demand for storage tends to double every 12-18 months, and adding in 80GB chunks at $18/GB is going to hurt. Sure, prices will drop, but not fast enough to be competitive with the multi-TB disks we're already seeing today.
Finally, suppose instead that you buy 2 of these SSDs to act as log devices and then buy 4 1TB 7200rpm SAS disks for $350 each. You've spent $4200 and you've gotten 24000 IOPS. That's $0.18/IOPS or $0.48/GB, and you've actually spent much less in absolute terms as well. You're still spending only a tiny fraction in power and cooling of what you were spending on the original all-disk solution, and you've got twice as much total storage capacity. Best of all, you can now grow your storage in two dimensions, not just along a line fixed by the disk vendors. Need more IOPS? Add another SSD or two. Need more capacity or streaming bandwidth? Add some more rotating rust.
This approach gives you the best of all worlds, something you can't get by blindly replacing all your disks with SSDs. In other words, you get to pick the spot along the performance/cost/capacity curve that's right for your application. Using only SSDs, only slow disks, or only expensive disks doesn't do that. Upon a moment's though, this should be obvious: when your computer needs to perform better, adding DRAM is usually the best way to make that happen. When it needs to store more data, adding disks is the way to go. You don't add disks to improve performance (one hopes... if you need to do that, your storage vendor is probably taking you to strip clubs) and you don't add DRAM to increase storage capacity. This is no different. Flash occupies an intermediate spot in the memory hierarchy and has to be thought of that way. It's exciting to see the prices fall and capacities rise like they have, but I don't think a lot of people really understand yet just how SSDs are going to change things.
Well I just checked and a hot disk in my SAN has done a bit over 15M 128K writes in the last 3 weeks so about 1.92TB in 21 days or close to 100GB per day. I have replaced 3 drives out of 150 in the last 2.5 years (well 5 total but 2 were precautionary from the SAN vendor when trying to troubleshoot another issue). This is a pretty lightly utilized SAN, we need it more for capacity then pure I/O. I can see a busy installation doing 10x what we do without even pushing the same hardware to its limit.
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
From the techreport article :
Is it time to look at connecting these chips direct to the motherboard ? Avoiding the added complexity of driving what is essentially a block of memory via a serial interface designed to control spinning discs. If the SLC memory chips were mapped into the main memory address space, it should be possible to make them look like a 32G or 64G (NV)RAM drive on a Unix/Linux system. Mount '/' and '/boot' on the (NV)RAM drive and install the OS on it. Presto - very fast boot and load times. You can still use traditional spinning disc(s) for large data, mounted as separate data partitions.
It would need some thought as to which parts of the filesystem went on spinning disc and which parts went on the (NV)RAM partition. But that is why Unix/Linux has all of the tools for mounting different parts of the filesystem on different partitions. Back in the olden days, most systems had a combination of small fast(ish) discs and big(ish) slow discs, and tweaking fstab to mount different parts of the filesystem on different discs was a standard part of the install process. Most desktop systems now have one huge disc, and the standard Linux install dumps everything on one big '/' partition, but all the tools for optimizing the partition layout are still there.
How about an ultra quiet desktop workstation with no moving parts, the OS installed on (NV)RAM disc, and user data dragged across the network from a fileserver (e.g NFS mounted /home).