Glonoinha's Journal: RamDrive!

Ok - the quest for higher knowledge and enlightenment continues. Not.

Actually I'm just looking to attain nerdvana - the fastest computing environment I can possibly manage.

This week Santa brought me exactly what I was hoping for ... six 512M DDR pc3200 sticks of memory for my mini-cluster. Two of the machines got bumped to 2G, one got a bump to 1.25G and the designated file server stayed at a pitiful 512M. Now WTF do I do with it (see previous question)?

RamDrive, of course. I downloaded the demo version of SuperSpeed's ramdrive package - it installs at the driver level and looks like just another piece of hardware. 45 Day trial period ought to be plenty. I installed it on the two machines with 2G apiece and broke out the stopwatches - figured that maybe with solid state drives I could finally realize the full potential of the gigabit backbone I put in last month. Two machines get configured, each with a 1.75G ramdrive, and out come the benchmark tests.

Initial feedback : positive, but not unGodly.
When staying on the same machine, copying files of any size from a hard drive to the SSD is still limited by the read throughput of the hard drive - in this case about 35-38 megabytes per second. This isn't bad, and is a HELL of a lot faster than when copying a file to a new place on the same hard drive.
When copying a single big file (800M) to another place on the same SSD things start to heat up - 220 - 240 megabyte/sec range. Given it is reading and writing on the same bus (full duplex) that's close to 500MB/s bandwidth on the memory bus.
I still haven't come up with anything that can benchmark peak read or write throughput, although Nero is reporting my hard drive at 55MB/s and the ramdrive at 1,392MB/s ... for what that's worth.
Moving a big file (1.5G) over the network from one ramdrive to another was a little disappointing - 53MB/s. Over a GigE back end (Intel integrated NIC's on PowerEdge400sc machines, Netgear 5 port GigE switch, hand crimped (by me) cables using Cat5 hardware.) Using regular file copy (command line in a DOS window to avoid any GUI overhead.) I was expecting quite a bit better than this - not sure what to think, where the bottleneck is...

On one machine I have moved the IE (yea, I know ...) temporary files onto the ramdrive and IE seems snappier, faster to render the pages (note that I'm on a 3Mb/s connection so ...) On pages with zillions of small files / graphic files it seems to make a difference. Nothing notable via numion.com however.

Stay tuned for more insights ... same Bat Time ... same Bat Channel.

cool stuff

[Mantorp]

do keep us posted, would be interesting to see if you find out where the bottleneck is.

RamDrive Benchmark Updates

[Glonoinha]

Ok, came up with a slew of benchmarks - I will first describe the tests, then give the numbers, then try to interpret the data.

Four machines, all Dell PowerEdge400sc boxes with P4 2.8GHz hyperthreaded CPUs, running XP Pro sp1, on a Gigabit backbone, Netgear 5 port switch.
Integrated Intel Gigabit NIC's on the motherboard.
Two of the machines have four 512M pc3200 sticks of memory in them (possibly running at pc2700 - the 400sc is real picky about memory timings - but it is running dual channel for what that'

Dynamic data cache

[EuropeUnited]

You wont have much use for the dedicated RAM-drives. Ever since at least the NT5-kernel (perhaps earlier in the NT-versions), Windows uses a dynamic cache of the same sort as linux has. All data that comes in or out of a block device goes thru the cache and stays there until it's pushed out by applications or new data.

If you have a single 1.65gb file (on a 2gb mem machine) that you want to serve repeatedly over the network, you'll probably not notice much of a difference between serving it from the harddri

Re:Dynamic data cache

[Glonoinha]

It was more a benchmarking tool than for any practical purpose.
Purely an experiment in computer science, I assure you - I have a GigE backbone and by golly I wanted to saturate it.

I have come to the realization that ramdrives are a pipe dream. You can generally make a ramdrive big enough to benchmark, let you know that your dreams are well justified and just how fast your real apps would go if you could put them entirely in ram - but you generally can't make the ramdrive big enough to actually do so.

Re:Dynamic data cache

[EuropeUnited]

Well, I think that you still can buy "true" ram-drives with a 3.5" form factor and a scsi- or ide-connector. Since they come with built in battery backup, they can be used just like an ordinary discdrive, with the difference of being able to transfer data at wire-speed and having a access time of, say, 0.01ms...

Of course, they do come at a price premium over regular magnetic drives. =)

Re:Dynamic data cache

[Glonoinha]

HyperOs HyperDrive III (16GB, ATA100) [hyperos2002.com] (scroll to the bottom) - I have the sales number for these guys in my cell phone [phone]). I called them before Christmas with high hopes but the thing isn't available in the US yet. Eight slots for memory (will take up to 2G DIMMs), battery backup, ATA100 interface so it is very portable between systems. Hella expensive though - last I checked 2G SDRAM chips were in the $600-$900 range (apiece, and it has room for 8) plus the $600 or so for the unit itself. W

Re:Dynamic data cache

[EuropeUnited]

Naa, I don't have any current knowledge about that kind of ram-drives. I just remember them from junior high, when I spent two weeks of "try-out-a-possible-future-career" at the IT-support in the Saab factory in Trollhättan, Sweden.

This wasn't all to long ago, I think it was in -98, but they were running 386-class machines as local process/robot controllers and such. Due to the vibrations, regular harddrives didn't last more than a few weeks/months, so they had ~$4000 ram-drives in them instead. I w