[Sheepdog] object sizes

Christoph Hellwig hch at infradead.org
Sun Nov 13 12:10:53 CET 2011


On Sat, Nov 12, 2011 at 05:02:50PM +0800, Liu Yuan wrote:
> I can not give a simple answer yet before I finish it. For your
> description about large data size, I have no doubt that it will benefit
> rotational disk utilizing the underlying extent-like feature and
> neutralize the seek overhead, but how about network performance for big
> data transfer? It seems to me that we trade off the fragmentation
> against wasteful bandwidth. A single 128MB object will fill the 1000Mb
> nic to the full, so parallel data transfer seems very restricted. Being
> that said, though I don't have much thought on this issue and my
> networking knowledge is not qualified to make the pertinent remark.

We don't nessecarily have to transfer all 128MB in a single I/O, just
make sure we allocate large sequential areas.  Many I/Os will be smaller
anyway, so it won't matter, but at least we have a way to fill the disk
queue with I/Os that avoid seeking if we want to (for large, streaming
I/O).




More information about the sheepdog mailing list