[PATCH] speedup sequential file reading

[apache2]

This patch changes the pread-like function for FILE* streams so
they use ftell() before calling fseeko(), and only calls fseeko()
when we are not at the offset we wanted to be at.

It may seem a bit backward to add yet another system call to the fast path here,
but for the zfs.pk pickle I added earlier it almost halves the runtime (good).

The problem seems to be that glibc discards its read buffer when you all fseek, 
regardless of whether it's practically a no-op, so this might not be that great 
on other libcs/platforms.

There are many things we could do to be more clever about it; like perhaps we 
could fread() our way to (offset) instead of fseek, and thus avoid clearing our 
buffers.
But I kind of like this patch; it doesn't add a whole lot of complexity, and it 
solves the problem (for me). :-)

I have attached the before-and-after 'perf' histograms, I apologize for the 
slightly different formats.

diff --git a/libpoke/ios-dev-file.c b/libpoke/ios-dev-file.c
index 66b451a6..1ab310c4 100644
--- a/libpoke/ios-dev-file.c
+++ b/libpoke/ios-dev-file.c
@@ -230,8 +230,9 @@ ios_dev_file_pread (void *iod, void *buf, size_t count, 
ios_dev_off offset)
 
   /* We are using FILE* for buffering, rather than low-level fd, so we
      have to fake low-level pread by using fseeko.  */
-  if (fseeko (fio->file, offset, SEEK_SET) == -1)
-    return IOD_EOF;
+  if (ftell(fio->file) != offset)
+    if (fseeko (fio->file, offset, SEEK_SET) == -1)
+      return IOD_EOF;
   ret = fread (buf, 1, count, fio->file);
 
   if (ferror (fio->file))

perf.hist.0
Description: Text document

perf.hist.1
Description: Text document