Slashdot Mirror
SoHo NAS With Good Network Throughput?
An anonymous reader writes "I work at a small business where we need to move around large datasets regularly (move onto test machine, test, move onto NAS for storage, move back to test machine, lather-rinse-repeat). The network is mostly OS X and Linux with one Windows machine (for compatibility testing). The size of our datasets is typically in the multiple GB, so network speed is as important as storage size. I'm looking for a preferably off-the shelf solution that can handle a significant portion of a GigE; maxing out at 6MB is useless. I've been looking at SoHo NAS's that support RAID such as Drobo, NetGear (formerly Infrant), and BuffaloTech (who unfortunately doesn't even list whether they support OS X). They all claim they come with a GigE interface, but what sort of network throughput can they really sustain? Most of the numbers I can find on the websites only talk about drive throughput, not network, so I'm hoping some of you with real-world experience can shed some light here."
If you want decent throughput build it yourself. Seriously. I have a coworker that bought 5 different NAS devices to do a bakeoff for a small skunkworks office and they all sucked for throughput. We ended up buying a $1K NAS that still wasn't great but sure beat all the SOHO ones. Numbers were ~8MB/s max on the fastest SOHO unit vs 25MB/s on the midrange one.
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
If you use a single disk NAS solution and you are doing sequential reads through your files and file system, your throughput can't be greater than the read/write speed of a single disk, which is no where near GigE (1000 Gbps is about 125 MB/second ignoring network protocol overhead). So you will need RAID (multiple disks) in your NAS, and you will want to use striped RAID (RAID 0) for performance. This means that you will not have any redundancy, unless you go with the very expensive striped mirror or mirrored stripes (1+0/0+1). RAID 5 gives you redundancy, and isn't bad for read, but will not be that great for writes.
As you compare/contrast NAS device performance, be sure that you understand the disk architecture in each case and see oranges to oranges comparisons (i.e, how does each one compare with the RAID architecture that you are interested in using - NAS devices that support RAID typically offer several RAID architectures). Also be sure that the numbers that you see are based on the kind of disk activity you will be using. It doesn't do much good to get a solution that is great at random small file reads (due to heavy use of cache and read-ahead) but ends up running out of steam when faced with steady sequential reads through the entire file system where cache is drained and read-ahead can't stay ahead.
Once you get past the NAS device's disk architecture, you should consider the file sharing protocol. Supposedly (I have no authoritative testing results) CIFS/SMB (Windows file sharing) has a 10% to 15% performance penalty compared to NFS (Unix file sharing). I have no idea how Apple's native file sharing protocol (AFP) compares, but (I think) OS X can do all three, so you have some freedom to select the best one for the devices that you are using. Of course, since there are multiple implementations of each file sharing protocol and the underlying TCP stacks, there are no hard and fast conclusions that you can draw about which specific implementation is better without testing. One vendor's NFS may suck, and hence another vendors good CIFS/SMB may beat its pants off, even if the NFS protocol is theoretically faster than the CIFS/SMB protocol.
Whichever file sharing protocol you choose, its very possible it will default to operation over TCP rather than UDP. If so, you should pay attention to how you tune your file sharing protocol READ/WRITE transaction sizes (if you can), and how you tune your TCP stack (windows sizes) to get the best performance possible. If you use an implementation over UDP, you still have to pay attention to how you set your READ/WRITE buffer sizes and how your system deals with IP fragmentation if the UDP PDU size exceeds what fits in a single IP packet due to the READ/WRITE sizes you set.
Finally, make sure that your network infrastructure is capable of supporting the data transfer rates you envision. Not all gigabit switches have full wire-speed non-blocking performance on all ports simultaneously, and the ones that do are very expensive. You don't necessarily need full non-blocking backplanes based on your scenario, but make sure that whatever switch you do use has enough backplane capacity to handle your file transfers and any other simultaneous activity you will have going through the same switch.
How many gigabytes are "multiple" gigabytes? Seriously, moving around five GB is much easier than 50 GB and enormously easier than 500 GB.
Another thing to consider: how many consumers are there? A "consumer" is any process that requests the data. If this post is a disguised version of "how do I serve all my DVD rips to all the computers in my house" then you probably won't ever have too many consumers to worry about. On the other hand, I work for an algorithmic trading company; we store enormous data sets (real-time market data) that range anywhere from a few hundred MB to upwards of 20 GB per day. The problem is that the traders are constantly doing analysis, so they may kick off hundreds of programs that each read several files at a time (in parallel via threads).
From what I've gathered, when such a high volume of data is requested from a network store, the problem isn't the network, it's the disks themselves. I.e., with a single sequential transfer, it's quite easy to max out your network connection: disk I/O will almost always be faster. But with multiple concurrent reads, the disks can't keep up. And note that this problem is compounded when using something like RAID5 or RAID6, because not only does your data have to be read, but the parity info as well.
So the object is to actually get many smaller disks, as opposed to fewer huge disks. The idea is to get the highest number of spindles as possible.
If, however, your needs are more modest (e.g. serving DVD rips to your household), then it's pretty easy (and IMO fun) to build your own NAS. Just get:
You might also want to purse the Ars Technica Forums. I've seen a number of informative NAS-related threads there.
One more note: lots of people jump immediately to the high performance, and high cost RAID controllers. I personally prefer Linux software RAID. I've had no problems with the software itself; my only problem is getting enough SATA ports. It's hard to find a non-server grade (i.e. cheap commodity) motherboard with more than six or eight SATA ports. It's even harder to find non-PCI SATA add-on cards. You don't want SATA on your PCI bus; maybe one disk is fine, but that bus is simply too slow for multiple modern SATA drives. It's not too hard to find two port PCI express SATA cards; but if you want to run a lot of disks, two ports/card isn't useful. I've only seen a couple of four-port non-RAID PCIe SATA cards. There's one eight port gem, but it requires PCI-X, which, again, is hard to find on non-server grade boards.
While it is true that the outside of the disk is spinning faster than the inner portion, in a modern HDD there are also several times more sectors in those outer rings. So while strictly speaking the read times might be faster, the seek times are not, and may even be slower. The sectors might even be interleaved, making any such comparison almost meaningless.
However, as you say, benchmarking is the only way to really tell. Highly recommended.
The only shops that actually look at cost/GB as a measuring stick are small shops, or shops with very specific needs.
Large corporations, government and high tech companies are usually more concerned with management costs, retention, migration and so forth.
This is simply not true. There are plenty of commodity storage requirements that do not require Fibre Channel or even NetApp level NAS. On the other end of the spectrum, cost/GB might not be a huge factor, but the cost of getting necessary IOPS is certainly a factor.
I work on Wall St. and we have multiple PB of storage. We have tons of EMC. However, things like the Sun X4500 and similar products from HP are changing the game. Couple that with being able to do 48 ports of line-rate 10GigE in a 1 RMU stackable, per priority pause coming into use, and Data Center Ethernet down the road and you have many reasons to seriously reconsider the scope of your fibre channel deployment.
"Nature doesn't care how smart you are. You can still be wrong." - Richard Feynman