On Wed, 13 Nov 2013 17:13:39 +0900 Christian Balzer <[email protected]> wrote: > On Wed, 13 Nov 2013 07:29:13 +0100 ml ml wrote: > > i have a simple 2 Node Setup. They are directly connected. We are > > now planning to implement SSD Disks. > > I now have the fear that the gigabit link between the two hosts will > > become a bottle neck. > It would be even with "normal" disks, certainly so if you have more > than one (RAID). > > Do you think bonding or tcp multipath will increase this? What > > would you recommend? > A google search finds the section of the DRBD user guide first, and > funnily enough a thread with some morsels of wisdom by yours truly as > well: > > http://lists.linbit.com/pipermail/drbd-user/2010-October/014858.html > > > Some mix between performance and redundancy is the goal. > > Has someone got real life production experience here? > Search and you will find many examples in the list archives and > elsewhere. > > Up to about 200MB/s replication speed and with a small budget, bonded > (directly connected) GbE links are fine. > > Once you need speeds over 250MB/s and/or fast I/O (transactions) I > would recommend directly connected Infiniband.
We are currently getting good results with two bonded direct-connected 10G links, 20Gbits seems almost fast enough for us:-) But I am planning on breaking that up to connect all three nodes with three 10G direct connections. And then I will either stay with 8.3 and distribute the drbd-volumes across the three nodes or go for drbd9 (or sheepdog or ceph) and do autobalancing/3-way-redundancy... If 10G (or infiniband) isn't in your budget, two bonded 1G connections are fine. Provided they are for drbd only. If you find the write-latency not good enough think about whether you need protocol C or if A or B are sufficient. Have fun, Arnold
signature.asc
Description: PGP signature
_______________________________________________ drbd-user mailing list [email protected] http://lists.linbit.com/mailman/listinfo/drbd-user
