bottlenecknetwork-attached-storagenetworking
I recently built a NAS box on top of ubuntu server. I have an average transfer of about 10mb/sec over gigabit wire to my computer. I was hoping you guys could give recommendations on how to pinpoint where the bottleneck is in this system. To my knowledge the router is only rated at 100mb/sec, and so are the nics (both server, and client), but shouldn't my files be transferring faster then 10mb/sec? I made sure to shutdown all other traffic when getting the 10mb/sec baseline.
Thanks,
Max.
First it is best to define the difference between a block device and filesystem. This is easier grasped if you are familiar with UNIX because it makes an objective distinction between the two things. Still the same applies to Windows.
/dev/sda for a disk or /dev/sda1 for a partition on that disk.mount /dev/sda1 /mnt/somepath.With those terms in mind it's then easier to see the distinction between the following.
I would suspect the RAID5 to be the main culprit. Every block write to your array could result in a reads followed by two writes - at best (if the extra block is already read into cache) it will be two writes. On top of that, if your transaction logs and data files are on the same array each write the database performs will result in two writes (one to the log and once to the data file) - while this will be the case on your home system too so it won't make much odds in your comparison on its own, it will compound any performance difference that there is otherwise.
Your home drive probably has its cache turned on, which will help by allowing it to reorder writes slightly to improve performance by reducing head movement. Turning on read-ahead may help you by reducing head contention with the write operations but that depends on the exact I/O pattern your query imposes.
RAID10 is usually recommended for databases rather than 5 for write performance reasons.
When you say "the same data" - is one system running a backed up and restored database from the other? If there is a difference in the history of the two databases it could be that the slower one needs to have its index stats updated (unexpected query speed problems can be a result of bas stats causing the query planner to chose a less optimal path than it otherwise would).
Also, do you have the same number of CPUs/cores on both environments? Some time ago I saw SQL Server 2000 take longer to run a disk-bottlenecked query slower when using two CPUs than when it was told just to use the one, I presume because it tried to split the load over two CPUs but this resulted in increasing I/O contention due to two threads reading together from different places on the disk (I've not seen this more recently, so it may have been a bug in that edition+SP that is long since fixed, but it might be worth looking in to - if you are giving explicit index hints that help in a low-core environment try removing them in case the query planner could make choices better suited to the extra cores, or try manually reoptimising the hints yourself).
Best Answer
Was this for transferring a single file? Also, is this performance consistent across multiple client machines? For testing, your best test is to take a very large file (500MB or so) and transfer it.
Here is how I would test it:
If all of the above yield 10Mb speeds, the issue is your router. If the performance is only slow when transferring to/from the NAS, then the issue is your Ubuntu setup (I'd guess network config, bad driver, etc). If the issue is only from your current test machine, but the borrowed one works fine, then it's likely a driver issue or config (e.g. duplex setting).
Hope this helps