On Fri, 23 Dec 2005, Burke, S (Stephen) wrote:
> > Even read-only access is problematic, because it means a file
> > cannot be
> > transparently moved to another file system or disk server, as
> > there may
> > be applications that have the file open for reading.
>
> Isn't that what pinning is supposed to solve?
I suppose it could, if it were taken to mean that! Pinning a file
means you will not have to wait for it to be staged in; it does not
have to mean the SE shall keep the TURL unchanged.
Furthermore, an SRM SE may have many disk servers: it would be a bit of
a nightmare to mount all of them on each WN. If one server goes down,
various programs on the WN may hang. A virtual file system spanning all
the physical file systems would be a solution, like dCache's PNFS; but
PNFS does _not_ allow one to just open()/read()/write()/close() files,
only directory operations are supported (again to facilitate and optimize
the implementation). It is true that an application (e.g. wrapper script)
could "reserve" a TURL before directly reading the file via NFS, but there
is a better solution: GFAL. Currently it does not yet allow one to open()
a "grid" file directly: one has to use gfal_open(); there has been work on
a version that also does the right thing on a plain open(), which would
allow applications to work unchanged, with the right LD_LIBRARY_PATH.
> > But now there are the concepts of tactical and strategic SEs:
>
> Are there? I know there was some talk about it in the glite architecture
> doc, but I could never really see the point as in practice you have a
> continuum and not just two types, and we already have the concept of
> "volatile" storage space - which none of the software uses! Anyway, I've
> never seen anything appear that looks like a tactical SE, either in LCG
> or glite.
Me neither, but that may change eventually...
|