On Wed, 2006-12-20 at 16:21 -0600, Chris Adams wrote: > I'm copying a big chunk of data to a single file on an NFS server (in > this case, using ntfsclone to back up a NTFS partition). Both the > client and the server are on gigabit ethernet, and both are running FC6. > > I see the disk LED come on solid for several seconds, then start > flickering, then go out for a second or two. If I watch the network > switch, it does the opposite: no traffic, then a burst of traffic. > Because of this, my total throughput is only about 200Mb/s. > > My guess would be that the NFS client is letting a bunch of data queue > up, and that it doesn't start sending it until the disk has almost > filled the kernel buffers, so then the reads from the local disk stop > until the NFS client has written enough data to the server. > > I also tried doing (the tr is in there to prevent a sparse file): > > $ dd if=/dev/zero bs=1024k count=5000 | tr '\0' '\1' > /mnt/test/testfile > > I can see the network start and stop on this. I do get about 320Mb/s, > but that's still a long way from what it should be. > > Is there anything I can adjust to smooth out the network flow (and keep > this running to fill the network as much as possible)? > -- > Chris Adams <cmadams@xxxxxxxxxx> > Systems and Network Administrator - HiWAAY Internet Services > I don't speak for anybody but myself - that's enough trouble. > Ethernet only works well when the network is utilized about 50%. The interface cards use various techniques to avoid overrunning the networks, and the systems also cannot sustain long runs of data at extremely high speeds due to processing overhead and other details. You may see less than you believe due to various bits of overhead. The headers, packetizing, routing info and so forth also eat into the total transfer rate, but are not counted if you are only looking at the data transfer rate. If you have heartbeat on going, that will be 800 to 1K bits each one or two seconds as well, along with the attendant data collisions and retransmissions. In addition, if the transmission is not a block transmission, but packetized, there may be handshaking with each packet which confirms correct reception and in essence a clear to send reaction, again adding collisions and processing and routing overhead. Hope this helps. As to speed ups: 1. check that your system and routers have heartbeat shut off (you really don't need it except for certain types of networking issues and resolution and tracking. 2. Check the transfer method (i.e. block transfer vs packets). Choose Block transfer if possible in your application. 3. Make sure your cables are correct length and that the connectors are clean and not corroded or poorly attached. 4. If you router is wireless capable, but you are not using it, turn off the wireless. If you cannot turn it off, then decrease the power so no one will be likely to piggyback on it. 5. Check your routing and cabling to make sure it has minimal router interfaces. i.e. direct connections if possible. Finally, make sure your system is not running some background task (such as yum updates) to eat up bandwidth. YMMV Regards, Les H
