I attempted to improve the bandwidth of a pipe by enlarging the kernelside
buffer if a writer offers more data than there's room. Currently a pipe
write is broken up into one-page chunks. The theory was that this would
reduce the number of context switches and system calls, and therefore improve
performance.
The results are quite surprising: bandwidth is reduced by about 100 MB/sec.
Here are some numbers:
2.2.14:
bw_pipe 353.97 MB/sec
bw_unix 139.80 MB/sec
lat_pipe 6.47 usec
lat_unix 13.24 usec
2.3.38:
bw_pipe 243.94 MB/sec
bw_unix 137.54 MB/sec
lat_pipe 6.69 usec
lat_unix 13.33 usec
2.3.38-fatpipe:
bw_pipe 159.11 MB/sec
bw_unix 137.39 MB/sec
lat_pipe 6.84 usec
lat_unix 12.83 usec
These are the microbenchmarks from lmbench-2alpha11. Each test was run
five times and the results averaged. The bandwidth tests do 64k writes,
the latency tests do 1-byte writes.
I don't know why 2.2.14 does better than 2.3.38, but I am guessing that
the -fatpipe patch makes things worse because now we are copying
64k of data in and out of the kernel on each cycle, which is slow.
Notice that AF_UNIX bandwidth in the unpatched kernel is much worse than
pipe bandwidth, but in the patched version they are almost the same. An
AF_UNIX socket has always consumed up to 128k of data in one write.
The patch is appended.
zw
--- linux-2.3.38/fs/fifo.c Sun Jan 9 00:29:32 2000
+++ linux-2.3.38-fatpipe/fs/fifo.c Sun Jan 9 17:52:05 2000
@@ -38,6 +38,7 @@
inode->i_pipe = info;
init_waitqueue_head(PIPE_WAIT(*inode));
+ PIPE_ORDER(*inode) = 0;
PIPE_BASE(*inode) = (char *) page;
PIPE_START(*inode) = PIPE_LEN(*inode) = 0;
PIPE_READERS(*inode) = PIPE_WRITERS(*inode) = 0;
--- linux-2.3.38/fs/pipe.c Sun Dec 5 08:42:03 1999
+++ linux-2.3.38-fatpipe/fs/pipe.c Sun Jan 9 18:33:18 2000
@@ -29,6 +29,9 @@
*
* Reads with count = 0 should always return 0.
* -- Julian Bradfield 1999-06-07.
+ *
+ * Enlarge pipe buffers if we're being fed data in bigger chunks.
+ * -- Zack Weinberg, 2000-01-09.
*/
/* Drop the inode semaphore and wait for a pipe event, atomically */
@@ -108,7 +111,7 @@
read += chars;
PIPE_START(*inode) += chars;
- PIPE_START(*inode) &= (PIPE_SIZE - 1);
+ PIPE_START(*inode) &= (PIPE_SIZE(*inode) - 1);
PIPE_LEN(*inode) -= chars;
count -= chars;
buf += chars;
@@ -167,6 +170,29 @@
if (!PIPE_READERS(*inode))
goto sigpipe;
+ /* If the buffer is empty and can't hold the entire write,
+ get a bigger buffer. (But don't go higher than PIPE_MAX.) */
+ if (PIPE_LEN(*inode) == 0 &&
+ count > PIPE_SIZE(*inode) && count <= PIPE_MAX) {
+ unsigned long new = 0;
+ unsigned int order;
+
+ order = PIPE_ORDER(*inode) + 1;
+ while ((1<<(order+PAGE_SHIFT)) < count)
+ order++;
+
+ while (order > PIPE_ORDER(*inode)
+ && !(new = __get_free_pages(GFP_USER, order)))
+ order--;
+ if (new) {
+ free_pages((unsigned long) PIPE_BASE(*inode),
+ PIPE_ORDER(*inode));
+ PIPE_ORDER(*inode) = order;
+ PIPE_BASE(*inode) = (char *) new;
+ PIPE_START(*inode) = 0;
+ }
+ }
+
/* If count <= PIPE_BUF, we have to make it atomic. */
free = (count <= PIPE_BUF ? count : 1);
@@ -390,7 +416,7 @@
if (!PIPE_READERS(*inode) && !PIPE_WRITERS(*inode)) {
struct pipe_inode_info *info = inode->i_pipe;
inode->i_pipe = NULL;
- free_page((unsigned long) info->base);
+ free_pages((unsigned long) info->base, info->order);
kfree(info);
} else {
wake_up_interruptible(PIPE_WAIT(*inode));
@@ -579,6 +605,7 @@
inode->i_op = &pipe_inode_operations;
init_waitqueue_head(PIPE_WAIT(*inode));
+ PIPE_ORDER(*inode) = 0;
PIPE_BASE(*inode) = (char *) page;
PIPE_START(*inode) = PIPE_LEN(*inode) = 0;
PIPE_READERS(*inode) = PIPE_WRITERS(*inode) = 1;
--- linux-2.3.38/include/linux/pipe_fs_i.h Sun Nov 21 11:17:45 1999
+++ linux-2.3.38-fatpipe/include/linux/pipe_fs_i.h Sun Jan 9 18:18:19 2000
@@ -4,6 +4,7 @@
struct pipe_inode_info {
wait_queue_head_t wait;
char *base;
+ unsigned int order;
unsigned int start;
unsigned int readers;
unsigned int writers;
@@ -13,7 +14,11 @@
/* Differs from PIPE_BUF in that PIPE_SIZE is the length of the actual
memory allocation, whereas PIPE_BUF makes atomicity guarantees. */
-#define PIPE_SIZE PAGE_SIZE
+#define PIPE_ORDER(inode) ((inode).i_pipe->order)
+#define PIPE_SIZE(inode) (1 << (PIPE_ORDER(inode)+PAGE_SHIFT))
+
+/* Limit on the space in kernel buffers. */
+#define PIPE_MAX (128*1024)
#define PIPE_SEM(inode) (&(inode).i_sem)
#define PIPE_WAIT(inode) (&(inode).i_pipe->wait)
@@ -26,10 +31,10 @@
#define PIPE_WAITING_WRITERS(inode) ((inode).i_pipe->waiting_writers)
#define PIPE_EMPTY(inode) (PIPE_LEN(inode) == 0)
-#define PIPE_FULL(inode) (PIPE_LEN(inode) == PIPE_SIZE)
-#define PIPE_FREE(inode) (PIPE_SIZE - PIPE_LEN(inode))
-#define PIPE_END(inode) ((PIPE_START(inode) + PIPE_LEN(inode)) & (PIPE_SIZE-1))
-#define PIPE_MAX_RCHUNK(inode) (PIPE_SIZE - PIPE_START(inode))
-#define PIPE_MAX_WCHUNK(inode) (PIPE_SIZE - PIPE_END(inode))
+#define PIPE_FULL(inode) (PIPE_LEN(inode) == PIPE_SIZE(inode))
+#define PIPE_FREE(inode) (PIPE_SIZE(inode) - PIPE_LEN(inode))
+#define PIPE_END(inode) ((PIPE_START(inode) + PIPE_LEN(inode)) & (PIPE_SIZE(inode)-1))
+#define PIPE_MAX_RCHUNK(inode) (PIPE_SIZE(inode) - PIPE_START(inode))
+#define PIPE_MAX_WCHUNK(inode) (PIPE_SIZE(inode) - PIPE_END(inode))
#endif
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Sat Jan 15 2000 - 21:00:15 EST