On Wed, May 27, 2015 at 04:24:29PM +0800, yuyang wrote: > Hello everyone, > > Recently we need copy one VDI from sheepdog to another storage > system(such as sheepdog or RAID). There are 2 ways to choose, > 1. dog vdi r | dog vdi w,(only for sheepdog, or temp file is needed) > 2. lib client, which is on the way. > > And then I did a R/W IO performance comparison. > Now I want to share the results. > > In this test, we read from sheepdog and write to the same sheepdog. Or we > just call > it "copy task". For the sake of simplicity, the sheepdog only has one node. > > The first step is to prepare the source VDI. The size of source VDI is 50GB, > and the data is 10GB/25GB/40GB(we do three tests). > > The task by lib client is done by the following code: > > /****** copy.c *********/ > #include <stdio.h> > #include <stdlib.h> > #include <sheepdog sheepdog.h=""> > > #define BUFSIZE (1024*1024*128) /**128M**/ > > void copy(struct sd_cluster *csrc, char *srcname, > struct sd_cluster *cdst, char *dstname) > { > int ret = 0; > struct sd_vdi * srcv = sd_vdi_open(csrc, srcname, 0, NULL); > if (!srcv) { > fprintf(stderr, "open %s failed\n", srcname); > return ; > } > uint64_t size = srcv->inode->vdi_size; > ret = sd_vdi_create(cdst, dstname, size, 0); > if (ret != SD_RES_SUCCESS) { > fprintf(stderr, "create failed %s\n", sd_error2str(ret)); > return; > } > struct sd_vdi * dstv = sd_vdi_open(cdst, dstname, 0, NULL); > if (!dstv) { > fprintf(stderr, "open %s failed\n", dstname); > return ; > } > char *buf = malloc(BUFSIZE); > if ( !buf ) { > printf("OOM\n"); > return ; > } > uint64_t offset = 0, left = size, onew; > while(offset < size) { > left = size - offset; > onew = left > BUFSIZE ? BUFSIZE : left; > memset(buf, 0, BUFSIZE); > > ret = sd_vdi_read(srcv, buf, onew, offset); > if (ret != SD_RES_SUCCESS) { > fprintf(stderr, " read error: offset=%"PRIx64, offset); > return; > } > ret = sd_vdi_write(dstv, buf, onew, offset); > if (ret != SD_RES_SUCCESS) { > fprintf(stderr, " write error: offset=%"PRIx64, offset); > return; > } > offset += onew; > } > free(buf); > } > > int main(int argc, char *argv[]) > { > char host1[256] = {}; > char host2[256] = {}; > memset(host1, 0, 256); > memset(host2, 0, 256); > if (argc != 5) { > printf("too small argument\n"); > printf("usage: copy srcIP srcvdiname dstIP > dstvdiname\n"); > return 0; > } > > strcpy(host1, argv[1]); > strcpy(host2, argv[3]); > > strcat(host1, ":7000"); > strcat(host2, ":7000"); > > struct sd_cluster *c1 = sd_connect(host1); > struct sd_cluster *c2 = sd_connect(host2); > > if (!c1 || !c2) { > printf("connect failed\n"); > return 0; > } > copy(c1, argv[2], c1, argv[4]); > return 0; > } > > compile: gcc copy.c -lpthread -lsheepdog > run it: > ./a.out 127.0.0.1 $vlibsrc 127.0.0.1 $vlibdst > > we record the runtime of the programme. > > And the copy task by dog cmd is done by the following script: > dog vdi create $vdogdst 50g > dog vdi read $vdogsrc | dog vdi write $vdogdst 0 50g > > we also record the runtime. > > Then we prepare 2 new source VDI in which there is 25GB(and then 40GB) > and repeat the steps above. > The result is as follows: > > data(GB) lib/dog time(s) > 10 lib 101 > 10 dog 239 > > 25 lib 128 > 25 dog 251 > > 40 lib 134 > 40 dog 249 > > As is shown in the table above, we can see that most of the time, > the lib client can speed up to more than 50% in VDI r/w.
Thanks for the share. I guess he performance gain is linear to the buffer size. Could you please add a dimension 'buffer size' such as following? 128M 256M 512M 1024M lib x y z .. dog x' y' z' .. Thanks, Yuan -- sheepdog mailing list sheepdog@lists.wpkg.org https://lists.wpkg.org/mailman/listinfo/sheepdog