On Thu, 5 May 2011, tsk wrote: > 2011/5/4 Wido den Hollander <wido@xxxxxxxxx>: > > Hi, > > > > On Wed, 2011-05-04 at 22:22 +0800, tsk wrote: > >> 2011/5/4 Wido den Hollander <wido@xxxxxxxxx>: > >> > Hi, > >> > > >> > On Wed, 2011-05-04 at 20:36 +0800, tsk wrote: > >> >> Hi folks, > >> >> > >> >> Is there any document for libceph, librbd and librados, and where > >> >> could I get some example for them? > >> > > >> > What kind of examples are you looking for? In the Ceph git repository > >> > there is: > >> > > >> > * rados.cc (The rados tool) > >> > * testradospp.cpp > >> > * testrbdpp.cpp > >> > >> Thx! > >> I have run over these test code yet, I think they are useful for me. > >> > >> Hmm, Is there any simple application example for these library? > > > > What kind of example are you looking for? I think everything you are > > searching for is in the testradospp/testrbdpp files, those really show > > what you can do. > > > > Could you explain what you are trying to achieve/write? > > > > Wido > > I have an application which is decided to be achieved by ceph, which > is the most suitable. But there are some requirements I don't know how > to satisfy with the ceph libs: > > There will be almost 10000 big file, each may be 10G~500G. Kernel > client will not be used, instead, a user space process will handle > read and write requests for each file via using the libs of ceph. > Because this will bypass the kernel VFS, and should get better > performance. > > Problem I want to issue: > 1. I need to set size of replicas for each file. e.g 1 replica for A > file, 2 for B file, 3 for C file.etc. > I think I should create several pools, each with a different replica > size. File for different replica size should be drop into the > corresponding pool. Is that workable? Yep. There is a per-file ioctl you can use to set the pool after creating the file but before writing any data to it, or there is a directory ioctl that sets the default layout/pool for new files created in that subtree. There should be equivalent libceph calls (if they aren't there yet, they are easily added in to complete the interface). > 2. Snapshot should be able to create, delete, list, rollback for each file. > Which APIs should I use in libceph, librbd and librados? The file system doesn't have a built-in 'rollback' operation. You currently need to recopy the old data over the new, unfortunately. It is possible to make this efficient (at least for individual files), but we haven't done it yet. > 3. The user space process and the cosd process will run on the same > host. To lower the network utilization, force-feeding will be used. I > guess reading performance may be better when do this, is it right? I > need to confirm this. How to configure the force-feeding optionÿÿ In > the libs or the CRUSH map? There is a 'preferred' field you can set via the libceph interface. There is also a get_local_osd() that tells you an osd on the local node (if there is one). sage
