Hi Jan, > why is PVFS2 not what I'm looking for? As far as I understand it is > possible with PVFS2 but not optimal in case of serial and small-file > workloads. What would be better for database and an apache server?
That would be hard to say since it depends on what kind of files you are serving off apache and what kind of db app you are running. Regardless, I am inclined to believe that most of the I/O operations from apache or databases tend to be smaller than what PVFS2 strives to optimize for. > PVFS2 (along with GFS, GPFS, Lustre, others) will do all the things that > > you describe in a single package (slightly reworded): > > - present lots of disks as one logical space > > - provide access to that logical space from many clients/nodes > > - coordinate access to provide consistency during concurrent access > > > > So you don't need to worry about using one thing to organize the disks, > > another thing to get the access to them, and a final thing to create the > > file system. iSCSI and/or FC may come up when fault tolerance comes > > into play, as described next. > > > > PVFS2 will also run in a fault tolerant configuration, but to do that > > multiple servers need access to the same storage. Usually that requires > > purchasing some FC or iSCSI hardware and dual-attaching to a couple of > > nodes. Most cluster file systems make this requirement for fault > > tolerance. Software redundancy (with performance) in this environment > > is very hard, and as far as I know no one is giving away an > > implementation. > > > you lost me here, I thought the multible servers already had access to the > same storage (the shared PVFS2 filesystem) via the network card. > why would I need iSCSI hardware can't I simulate it with something like > http://sourceforge.net/projects/iscsitarget/ > it is still not clear to me how PVFS2 is fault tolerant? I thought the > filesystem stays available but the data on the failed node is unavailable? > or is the data also mirrord on an other node? The multiple servers together stitch the PVFS2 file-system namespace to a bunch of clients (Typically, they do so by utilizing available disk-based file-systems attached to the nodes on which they run) (Note that servers themselves do not see the shared PVFS2 file-system unless of course you mount them explicitly on all server nodes.) However, in order to make PVFS2 fault tolerant, all servers need to be able to see (possibly all) other servers' storage spaces in order to serve data or meta-data requests. One way to do this would be to mirror any updates to the storage space on a set of servers which would be prohibitively expensive and would come at the expense of performance. Another way to do this would be to make all servers use a shared-disk infrastructure (made available using either iSCSI or FC) formatted as a regular file-system allowing multi-node read/write accesses (such as GFS). Thanks, Murali > > greez, > Jan > > > PVFS2 is different from most cluster file systems in that we don't do > > client-side caching or locking. This is good for fault tolerance > > because it simplifies failover scenarios, but it is bad for serial and > > small-file workloads because operations always go over the wire, and > > that can be slow. > > > > Hope that helps, > > > > Rob > > > > Jan Lenaerts wrote: > > > hellow, > > > > > > The past hour I've been reading about PVFS2, but it's still not clear to > > > me, if PVFS2 can do what I want or not. > > > Let me explain: > > > > > > I have 4 node's in a high availibility cluster, I want to bundle a part > > > of diskspace from each node, and use that as a shared filesystem for my > > > nodes. > > > so I need: > > > > > > * something that can bundle the diskspace(PVFS2 maybe? ) > > > * something to share the disk to each node(iSCSI target & > > > iSCSI_initiator) > > > * a FS that won't get corrupt when multible nodes read or write on > > > it(GFS or OCFS2) > > > > > > - But PVFS2 is already a filesystem is it possible to format GFS on top > > > of that? > > > - PVFS2 is at filesystem level, is it possible to share it on > > > block_device level to each node, as with iSCSI? > > > - does it preform good as FS for apache, mysql etc... ? > > > - If one node (dataserver fails) is the data somewhere else available? I > > > guess the data on that node is then unavailable? witch is unacceptable > > > in my case. Is there a way to do this? > > > - if PVFS2 isn't what I need, Then what is out there to become complete > > > fault redundant storage, something like > > > http://www.capricorn-tech.com/products.html or what I want > > > - How can I achieve this, is there maybe a guide that explains a setup > > > like I would like? > > > > > > Thanks in advance, > > > Jan Lenaerts > > > _______________________________________________ PVFS2-users mailing list [email protected] http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
