Acard ANS-9010B 5.25-inch SATA x1-to-6-slot DDR2 RAM Disk (Deactivated) | Hacker News Reader

Acard ANS-9010B 5.25-inch SATA x1-to-6-slot DDR2 RAM Disk (Deactivated)

scsi4me.com

33 pointsthunderbong43 comments

If you want a ramdisk maybe look into these: https://ddramdisk.store/2023/02/15/the-ddr4-pcie-x8-lady-suc...

I'm not in anyway connected to this, found it after looking for something modern after seeing a YouTube video about old ramdisks.

Didn't there used to be a version of this that was RAM plus a mechanical hard disk it would get dumped to when powered down? ("hybrid hard disk" seems to mean SSD+Mechanical now.)

I don't know. Only ever heard about the OP's thing with compact flash. But since cf is IDE I guess you can use spinning rust as well.

Powering a mechanical hard drive on an emergency battery seems hard. I know RAID cards have battery backed write caches that either store in RAM or flush to flash storage on power outage, and then rewrite when power is back on.

I can understand why, historically, it made sense to have something that looks like a hardware disk but uses RAM. But today, what's the use case for this, versus just adding the RAM to the system directly and using a software RAM disk? That avoids the disk protocol entirely, and should provide substantially better performance.

I believe this is battery backed, so it doesn't lose data when powered off. With that said, mmap hard disk file would get similar performance (minus fsync overhead to the persistent disk backing store).

Any idea how this compares to enabling write caching on a drive or on your database layer, if you're using one? I mean other than the catastrophic effects of a sudden power outage.

I'd think the main benefit these days might be cost savings from using RAM that's a generation or two older, but in general I think the use case for these hardware RAM disks is probably pretty slim nowadays.

Even with older RAM, if you need a large and fast storage, few solutions can compete with RAM. Especially when it comes to random R/W. You can get 4.4GBps random read from a DDR3 drive (in parents link) while the latest Gen5 NVME drives tend to cap out at ~90MBps for such loads (worst kind of loads). In addition these can be installed into any system, no need for a special server MOBO with enough RAM slots.

What poor NVMe are you buying that gets 90MB/s? Even the Gen3 Kingston 256MB drive I have does well over 500MB/s in my iozone tests on a Dell Precision 3420. The only time you need mobo support is if you want to boot off NVMe but every PCI-compliant OS should be able to at least see it.

> the latest Gen5 NVME drives tend to cap out at ~90MBps for such loads (worst kind of loads)

Possibly scattered 1-byte writes rather than sequential 4096-byte page writes.

Benchmark in English here: https://www.storagereview.com/news/gigabyte-aorus-10000-gen5...

Would it be better to user DDR3, or anything you can buy for cheap on ebay? You won't use the full speed of the RAM anyway, so why use those older RAM modules that would go to waste anyway.

Its a russian website and afaik everything was always "out of stock". "prices" wouldnt even cover cost of FPGA fast enough. Oh and they already "made" DDR3 board that looks the same, but never sold anything https://web.archive.org/web/20210714152011/https://ddramdisk...

There was a time when 1GB of memory cost maybe $50 but you had to pay a sizeable premium to buy a server capable of handling more than 4GB. Sites had racks full of 4GB servers acting as memcached boxes. It's a lot easier to buy servers that can hold 256gb or more of ram, so your expenditures are more balanced between ram and other stuff, and you are still using commodity ram.

> There was a time when 1GB of memory cost maybe $50 but you had to pay a sizeable premium to buy a server capable of handling more than 4GB.

Looking at the memory price timeline https://jcmit.net/memoryprice.htm it was middle 2007 when 1GB memory hit $50 (row 2007.33 it's a very big drop in $/memory as you can see in the second column). A contemporary server motherboard would be the Supermicro X7DCL https://www.supermicro.com/manuals/motherboard/5100/MNL-0957... with six slots of DDR2 memory capable of handling 48GB RAM https://ark.intel.com/content/www/us/en/ark/products/33142/i... but even with cheap memory it was easily doing 12-24GB RAM.

Looking at it from another angle, https://github.com/memcached/memcached/wiki/ReleaseNotes150 "Memcached was introduced in 2003" which is a bit after Socket 604 and looking at those boards https://www.supermicro.com/products/motherboard/archive/?mlg... it does not seem like at that era 8GB maximum would be uncommon (8GB instead of 4GB because we want to use cheaper than top modules), in fact very few motherboards top out at 8GB, most could do a lot more. This is to say, there doesn't seem to be a period of history where memcached existed but commodity servers handling more than 8GB didn't exist.

Yet another angle, I upgraded from the Panasonic CF-Y5 to a ThinkPad T400s because the CF-Y5 topped out at 2GB and that was in 2009 (I checked my records, 2009 August 11) and so servers not handling vastly more is extremely unlikely. At this time the company I worked with certainly had dedicated memcached boxes yes, but a lot more than 4GB RAM, can't remember exactly how much.

I would cautiously say you misremember your numbers.

> with six slots of DDR2 memory capable of handling 48GB RAM

1) that was a premium motherboard, 2) 16 ram sticks were way more expensive per GB than smaller sticks. I think you have the timeframe about right (2008?), but where I worked we had 1000s of servers built on PC motherboards (cheap) with 4 slots holding 1GB each. 4GB sticks existed but cost a lot more per GB and were hard to find, and our motherboards needed some kind of bios change to use them. We did manage to upgrade a few machines to 16gb and they were useful.

I went looking for something like this for NVMe drives and there's surprisingly little available. External (USB3) adapters for NVMe drives can be had for under $20, so I'm surprised there aren't internal 2.5'' / 3.5'' options available (there are PCIe cards, but on a mini-itx board there are no extra slots besides the GPU).

Why spend the money for a super fast direct pcie ssd and then pay more money to cripple it down to sata speeds when you could just buy a sata ssd or msata+adapter way cheaper?

Also it's a complex adapter, since it has to be a PCI-e host. Generally SATA->x (USB, SD card, etc) adapters can be done with much simpler app-specific microcontrollers with IP blocks sprinkled in.

One reason is SATA SSDs are more of expensive than their m2 counterparts - at least they were here in the UK when I recently purchased an SSD.

Another reason could be size - in a 5.25" form factor, you could potentially "stack" several m2 SSDs in blade-like fashion. I haven't seen such a thing though, and heat might be an issue.

2.5” 2T Samsung ssd on Amazon is $135 and a 2T nvme is slightly cheaper at $125. That $10 in savings is going to totally be eaten in adapters though.

So there are 5.25” multiple m2 adapters out there like

https://www.icydock.com/goods.php?id=255

That one would require a sas controller to connect to.

You generally don’t see a tighter system like blades due to needing a bifurcated pcie slot support, which at that point of server hardware, your are generally not using adapters like this.

That enclosure isn't m.2, that's 2.5" NVMe. Like a SATA drive but with more pins stuffed in the plug. Though I guess you could put extra adapters inside it.

Is the ability to bifurcate a PCIe slot (and avoid extra controllers) that rare? That's a shame.

Whoops. I linked the wrong model. Middle of that page has a section on all the different options for the drive bay. A few that look very similar are m2 nvme.

I have a https://www.supermicro.com/en/products/motherboard/X10DRH-iT and they don’t support bifurcated pcie. It’s a shame cause I’ve extended the life of the servers by using pcie adapters to add high speed io layers and I’m limited to one device per port. I’d love to double that density.

SATA SSDs are a little bit cheaper here.

Hot swap bays are usually set up for 2.5" drives. You can fit 4-8 into a single 5.25" slot. But the ones designed for performance will prefer U.2 drives and hook them up to PCIe.

https://www.icydock.com/goods.php?id=321

I agree it's a pretty niche requirement. But I have 8x 1Tb NVMe drives after pulling them out of a server and nothing to use them for :)

This is exactly what CXL is for.

Gonna take quite a while for such raw versions to emerge. Gonna be stupid expensive enterprise drives for years. But really it could be just a chip with some ram controllers & a CXL port.

Alas the fpga market is a lumbering hulking ancient shit show. Lattice put out a solid mid range part but the whole rest of the market is dead, unless you want chips that cost as much as cars for milspec radars or whatnot, then there's great shit. What a hellhole, such a core base for computing, unserved.

There's one for sale on eBay if you're looking to buy?

https://www.ebay.com/itm/125610100927

"Data transfer rate up to 200MB/sec"

So it's not even as fast as a modern SATA SSD. Oof.

Depends on the modern SATA SSD, QLC SSDs are quite a bit slower than 200 MB/s for writing (something like a Samsung 870 QVO manages about 150 MB/s real-world, cheaper ones are probably even slower), so they're actually slower than current HDDs for sequential writes.

Interesting, before cheap ram disks maybe had a use case, but now motherboards have 2/3 m2's with raid options. Or even jbod ssd's for software style raid.

Uses ddr2 and ram was expensive when it came out, so can see it wasn't super popular. But neat idea.

https://techreport.com/review/16255/acards-ans-9010-serial-a...

There's still a use-case for them. High performance high reliability persistent messaging is one such use case where SSDs do not have the required endurance. Solace builds their hardware messaging appliances with the Assured Delivery Blade that consists of an FPGA based board with supercap backed DRAM that dumps to flash on power failure and a couple of 10Gbps ports to mirror to another system. The end result is extremely low latency acknowledgement of non-volatile storage of persistent messages with unlimited endurance.

With that being a 5.25, year and brand I was thinking desktop usage. For enterprise, I'm sure theres always a need for some.

Man, that is a blast from the past. I had an engineering sample of one of these in 2008 and was trying to get it to work with ECC RAM pulled out of a Sun box. What a blast from the past.

https://techreport.com/review/16255/acards-ans-9010-serial-a...

Linus Tech Tips reviewed the Gigabyte i-RAM SATA drive back in year 2019. https://www.youtube.com/watch?v=bYbCYgYZVT8 , https://en.wikipedia.org/wiki/I-RAM

I think if you're doing LOTS of writes to something you don't care a lot about - a ramdisk might be a nice way to prevent you from killing your SSD.

Say, most things you put in /var

This particular unit might also work for a machine where you install a client OS on the main (ramdisk) HD and reboot, then wipe and do it again later.

>I think if you're doing LOTS of writes to something you don't care a lot about - a ramdisk might be a nice way to prevent you from killing your SSD.

>Say, most things you put in /var

Even a bottom of the barrel QLC SSD can handle 100GB of writes per day without problems. Therefore it's unlikely that moving stuff in /var (random log/temporary files) off the SSD is going to change its lifespan by much.

video editing?

If you really like video editing splurge and get a 4 TB drive with ~2.5 PBW rating for $250, saves you from needing to copy things to specific locations and at 1 TB per day every day you shouldn't have to start thinking about a replacement for 6 years.

That's enough to write a 100 Mbps high quality local copy of a 24/7 stream for the 6 years.

I have 4 of these each with 16GB ECC RAM. Used them as a ZFS cash. They run kinda hot.

The more I think about it the less sense a ram disk makes, I will admit it, I have abused them in the past, once for battlefield3 because the load times on my setup were awful and once for a company, they wanted a lookup using a specific database product that had no way to do in memory tables.

But resorting to a ramdisk means you have failed. most of the time the page cache offers the semantics you actually want, memory quick access backed up by persistent storage. You are saying, I want a filesystem api with ram speed and are willing to jump through the hoops required, manually syncing your storage at startup and shutdown, to reach that specification.

> most of the time the page cache offers the semantics you actually want, memory quick access backed up by persistent storage.

If you are thinking about a ram disk it is likely because you want guaranteed low latency. With a cache you get hosed by delays when pages are flushed. I remember having to run daemons just to keep caches warm by periodically accessing data that I might need later without having to wait for it.

RAM disks had their heyday in the 32 bit era, where computers could be limited to 4GB of RAM but if you wanted to you could get a much larger RAMdisk.

Reply on news.ycombinator.com