Browse Source

EDF reading emulation (iteration1)

Suren A. Chilingaryan 11 years ago
parent
commit
ce7895919b
3 changed files with 339 additions and 0 deletions
  1. 1 0
      .bzrignore
  2. 2 0
      CMakeLists.txt
  3. 336 0
      rndreader.c

+ 1 - 0
.bzrignore

@@ -5,3 +5,4 @@ config.h
 cmake_install.cmake
 fwbench
 seqreader
+rndreader

+ 2 - 0
CMakeLists.txt

@@ -35,6 +35,7 @@ include_directories(
 
 add_executable(fwbench fwbench.c)
 add_executable(seqreader seqreader.c)
+add_executable(rndreader rndreader.c)
 
 if (USE_UFO_GENERATOR)
     target_link_libraries(fwbench m ufo fastwriter ${GLIB2_LIBRARIES} ${GTHREAD2_LIBRARIES})
@@ -43,5 +44,6 @@ else (USE_UFO_GENERATOR)
 endif (USE_UFO_GENERATOR)
 
 target_link_libraries(seqreader aio)
+target_link_libraries(rndreader aio)
 
 configure_file(config.h.in ${CMAKE_CURRENT_BINARY_DIR}/config.h)

+ 336 - 0
rndreader.c

@@ -0,0 +1,336 @@
+#define _GNU_SOURCE
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <sys/time.h>
+#include <unistd.h>
+#include <dirent.h>
+#include <fcntl.h>
+#include <string.h>
+#include <errno.h>
+
+#include <libaio.h>
+
+#define FASTWRITER_SYNCIO_ALIGN 4096
+
+#define SYNC_MODE
+#define AIO_MODE 2
+#define EXTRA_BUFFERS 2
+#define WRITE_INTERVAL 1
+
+size_t SKIP = 1;
+size_t SEGMENT = 0;
+size_t LINE = 0;
+
+#define RAID_STRIP_SIZE 	256
+#define RAID_DISKS		8
+#define STRIPS_AT_ONCE		2
+
+#define MIN_BLOCK_SIZE (1024 * RAID_STRIP_SIZE)
+#define BLOCK_SIZE (1024 * RAID_STRIP_SIZE * RAID_DISKS * STRIPS_AT_ONCE)
+#define BUFSIZE (BLOCK_SIZE * (AIO_MODE + EXTRA_BUFFERS))
+
+#ifdef AIO_MODE
+# define SYNC_MODE
+#endif /* AIO_MODE */
+
+
+DIR *dir; 
+struct dirent *ent; 
+int flags = O_RDONLY|O_NOATIME|O_LARGEFILE|O_DIRECT;
+
+size_t run = 0; 	// file reading order
+size_t skip = 0; 	// file reading order
+
+int fd = -1;		// we are currently scheduling reads from this file
+size_t offset = 0;	// we are schedulling reads at this offset
+size_t size;		// size of block we are schedulling to read
+size_t file_size; 	// total size of current file
+size_t file_block_size; // access unit for current file
+
+size_t curf = 0, schedf = 0;
+size_t curio = 0, schedio = 0;
+
+#define MAXLEN 128
+int done[AIO_MODE + EXTRA_BUFFERS];
+int done_finish[AIO_MODE + EXTRA_BUFFERS];
+char done_file[AIO_MODE + EXTRA_BUFFERS][MAXLEN + 1];
+size_t done_offset[AIO_MODE + EXTRA_BUFFERS];
+size_t done_size[AIO_MODE + EXTRA_BUFFERS];
+int done_fd[AIO_MODE + EXTRA_BUFFERS];
+
+#define buf(io) (io%(AIO_MODE + EXTRA_BUFFERS))
+#define max(a,b) (((a)>(b))?(a):(b))
+#define min(a,b) (((a)<(b))?(a):(b))
+#define page(size) (((size/4096)+(size%4096?1:0))*4096)
+
+
+int next_file() {
+    size_t size;
+    struct stat st;
+    
+next:
+    while ((ent = readdir(dir))) {
+	skip += 1;
+	if ((((skip) - 1)%SKIP) != run) continue;
+	if (stat(ent->d_name, &st)) continue;
+	if (!S_ISREG(st.st_mode)) continue;
+	break;
+    }
+
+    if (ent) {
+	size = st.st_blksize;
+
+	int fd = open(ent->d_name, flags, 0);
+	if (fd < 0) goto next;
+
+	if (size < MIN_BLOCK_SIZE) size = BLOCK_SIZE;
+
+	if (size > BUFSIZE) {
+	    printf("Buffer too small\n");
+	    exit(1);
+	}
+	
+	file_block_size = size;
+	file_size = st.st_size;
+	offset = 0;
+	
+	return fd;
+
+    } else {
+	skip = 0; 
+	run += 1;
+	if (run < SKIP) {
+	    closedir(dir);
+	    dir = opendir(".");
+	    goto next;
+	}
+    }
+
+    return -1;
+}
+
+int next_block() {
+    int io;
+    size_t next_offset;
+    
+    if (fd >= 0) {
+	if (LINE) offset += SEGMENT;
+	else offset += size;
+
+	if (offset >= file_size) {
+	    fd = -1;
+	}
+    }
+
+    if (fd < 0) {
+	fd = next_file();
+	if (fd < 0) return -1;
+//	printf("open ===> %s (%i)\n", ent->d_name, fd);
+    }
+
+    if (LINE) size = LINE;
+    else size = file_block_size;
+
+    if ((offset + size) > file_size) size = file_size - offset;
+    
+    io = buf(schedio);
+    strncpy(done_file[io], ent->d_name, MAXLEN);
+    done_offset[io] = offset;
+    done_size[io] = size;
+    done_fd[io] = fd;
+    done[io] = 0;
+
+    if (LINE) next_offset = offset + SEGMENT;
+    else next_offset = offset + size;
+
+    if (next_offset >= file_size) done_finish[io] = 1;
+    else done_finish[io] = 0;
+
+    return 0;
+}
+
+
+
+int main(int argc, char *argv[]) {
+    int err;
+    int i, n, io;
+    long double mcoef = 1000000. / (1024 * 1024);
+
+    io_context_t aio;
+    struct iocb ios[AIO_MODE], *ioptr[AIO_MODE];
+
+    int events;
+    struct io_event ev[AIO_MODE];
+
+    int ready;
+    void *buffer;
+    struct timeval start, fstart, tv;
+
+    if (argc < 2) {
+	printf("Usage: %s <directory> [skip] [segment] [line]\n", argv[0]);
+	exit(0);
+    }
+    
+    chdir(argv[1]);
+    dir = opendir(".");
+    
+    if (argc > 2) SKIP = atoi(argv[2]);
+    if (argc > 3) SEGMENT = atoi(argv[3]);
+    if (argc > 4) LINE = atoi(argv[4]);
+
+    if (!SKIP) SKIP = 1;
+    
+    printf("%s: Skip %zu, Segment %zu, Line %zu\n", argv[1], SKIP, SEGMENT, LINE);
+
+    posix_memalign((void**)&buffer, FASTWRITER_SYNCIO_ALIGN, BUFSIZE);
+
+    memset(done, 0, sizeof(done));
+    memset(&aio, 0, sizeof(aio));
+    io_queue_init(AIO_MODE, &aio);
+
+    for (schedio = 0; schedio < AIO_MODE; schedio++) {
+	ioptr[schedio] = &ios[schedio];
+	memset(ioptr[schedio], 0, sizeof(struct iocb));
+	
+	err = next_block();
+	if (err) break;
+	
+	io_prep_pread(ioptr[schedio], fd, buffer + schedio * BLOCK_SIZE, page(size), offset);
+	io_set_callback(ioptr[schedio], (void*)(uintptr_t)schedio);
+//	printf("sched %zu: %zu (%zu %zu)\n", schedio, schedio, offset, size);
+    }
+
+    size_t us, fileus; 
+    size_t last_write = 0;
+//    size_t last_file_write = 0;
+//    size_t last_file_size = 0;
+    size_t cur_file_size = 0;
+    size_t total_size = 0;
+    size_t files = 0;
+
+    gettimeofday(&start, NULL);
+    gettimeofday(&fstart, NULL);
+
+    n = io_submit(aio, schedio, ioptr);
+    if (n != schedio) {
+        printf("Failed to submit initial AIO job, io_submit returned %i\n", err);
+        exit(-1);
+    }
+
+    curio = 0;
+    events = 0;
+
+
+    ready = 0;    
+
+    while ((err >= 0)||(curio != schedio)) {
+	io = buf(curio);
+	
+	if (!done[io]) {
+//		printf("%i,%i - %i [%i %i %i %i]\n", curio, schedio, events, done[0], done[1], done[2], done[3]);
+
+		if (curio < schedio) {
+		    n = io_getevents(aio, 1, AIO_MODE + EXTRA_BUFFERS - events, &ev[events], NULL);
+		    if (n < 0) {
+			printf("Error waiting for AIO (%i)\n", -err);
+			exit(-1);
+		    }
+		} else {
+		    n = 0;
+		}
+		
+		if ((!ready)&&(n > 1)) {
+		    printf("*** Multiple read requests (%i of %i) are finished simultaneously. It is either:\n", err, AIO_MODE);
+		    printf("      Small buffer size (%i KB)\n", BLOCK_SIZE/1024);
+		    printf("      More parallel AIOs (%i) than supported by kernel, try %i\n", AIO_MODE, AIO_MODE - err);
+		}
+		
+		for (i = 0; i < n; i++) {
+		    struct io_event *ep = &ev[events + i];
+		    int doneio = (uintptr_t)ep->data;
+		    io = buf(doneio);
+//		    printf("done %i: %lu %zu %zi\n", doneio, ep->res2, done_size[io], ep->res);
+	    	    if (ep->res2 || (ep->res < done_size[io])) {
+	    		printf("Error in async IO (ret: %li, ret size: %zi, expected %zu)\n", ep->res2, ep->res, page(done_size[io]));
+	    		exit(-1);
+	    	    }
+		    done[io] = 1;
+//		    printf("done (%i): %i\n", i, doneio);
+		}
+		
+		events += n;
+		
+		for (i = events - 1; (i >= 0)&&((schedio - curio) < (AIO_MODE + EXTRA_BUFFERS)); i--) {
+		    err = next_block();
+		    if (err) break;
+		    
+//		    printf("sched %i: %zu (%i %zu %zu)\n", i, schedio, fd, offset, size);
+		    struct iocb *newio = (struct iocb *)ev[i].obj;
+		    memset(newio, 0, sizeof(struct iocb));
+		    io_prep_pread(newio, fd, buffer + buf(schedio) * BLOCK_SIZE, page(size), offset);
+		    io_set_callback(newio, (void*)(uintptr_t)schedio);
+		    err = io_submit(aio, 1, &newio);
+		    if (err != 1) {
+			printf("Failed to submit AIO jobs %i\n", err);
+			exit(-1);
+		    }
+		    schedio++;
+		}
+		events = i + 1;
+		
+		if ((events)&&(!err)) {
+		    printf("*** Unprocessed events (%i), probably not enough buffer space...\n", events);
+//		    printf("      curio (%zu), schedio (%zu)\n", curio, schedio);
+		}
+
+		ready = 1;
+		continue;
+	    }
+
+	    io = buf(curio);
+
+	    cur_file_size += done_size[io];
+	    total_size += done_size[io];
+
+	    gettimeofday(&tv, NULL);
+	    us = (tv.tv_sec - start.tv_sec) * 1000000 + (tv.tv_usec - start.tv_usec);
+	    fileus = (tv.tv_sec - fstart.tv_sec) * 1000000 + (tv.tv_usec - fstart.tv_usec);
+	    
+	    if (done_finish[io]) {
+//		printf("closing %i ===> %s (%i)\n", io, done_file[io], done_fd[io]);
+		close(done_fd[io]);
+	        gettimeofday(&fstart, NULL);
+		files++;
+	    }
+	    
+	    if ((us - last_write) > WRITE_INTERVAL * 1000000) {
+		last_write = us;
+		printf("Read: %lu files (%lu GB) at %zu MB/s", files, total_size / 1024 / 1024 / 1024, (size_t)(mcoef * total_size / us));
+		
+		if (1) {//cur_file_size > BLOCK_SIZE) {
+	    	    printf(", Last: %s (%lu MB) at %zu MB/s\n", done_file[io], cur_file_size / 1024 / 1024, (size_t)(mcoef * cur_file_size / fileus));
+	    	} else {
+	    	    printf("\n");
+	    	}
+	    }
+
+	    if (done_finish[io]) {
+		cur_file_size = 0;
+	    }
+	    
+	    done[io] = 0;
+
+	    curio++;
+	}
+
+        gettimeofday(&tv, NULL);
+        us = (tv.tv_sec - start.tv_sec) * 1000000 + (tv.tv_usec - start.tv_usec);
+	printf("Total: %lu files (%lu GB) at %zu MB/s\n", files, total_size / 1024 / 1024 / 1024, (size_t)(mcoef * total_size / us));
+	
+	free(buffer);
+	closedir(dir);
+}