Re: sparse file performance (was Re: Is there a "make hole" (truncate in middle) syscall?)

[Andy Isaacson <adi@hexapodia.org>]

On Fri, Dec 05, 2003 at 01:12:21PM -0800, Linus Torvalds wrote:
> On Fri, 5 Dec 2003, Andy Isaacson wrote:
> > I got curious enough to run some tests, and was suprised at the results.
> > My machine (Athlon XP 2400+, 2030 MHz, 512 MB, KT400, 2.4.22) can read
> > out of buffer cache at 234 MB/s, and off of its IDE disk at 40 MB/s.
> > I'd assumed that read(2)ing a holey file would go faster than reading
> > out of buffer cache; in theory you could do it completely in L1 cache
> > (with a 4KB buffer, it's just a ton of syscalls, some page table
> > manipulation, and a bunch of memcpy() out of a single zero page).  But
> > it turns out that reading a hole is *slower* than reading data from
> > buffer cache, just 195 MB/s.
> 
> That's because we actually instantiate the page cache pages even for
> holes. We have to, or we'd have to special-case them no end (and quite
> frankly, "hole read performance" is not something worth special casing,
> since it just isn't done under any real load).
> 
> So reading a hole implies creating the page cache entry and _clearing_ it.
> For each page. So while you may read from the L1, you also have to do
> writeback of the _previous_ pages from the L1 into the L2 and eventually
> out to memory.
> 
> (And eventually the VM also has to get rid of the pages etc, of course).

Thanks for the explanation, Linus.

I modified my benchmark to use mmap(2) instead of read(2) and the
results are broadly comparable.  With a 10MB window, I get 331 MB/s
reading out of buffer cache and 185 MB/s reading a hole.  Reading a file
too large to cache is about the same (disk-limited) speed, 43 MB/s.

-andy