diff --git a/Makefile b/Makefile
index b23534e..3c4abeb 100644
--- a/Makefile
+++ b/Makefile
@@ -3,8 +3,9 @@ CC=			gcc
 # CFLAGS=		-g -Wall -Wno-unused-function -O2
 CFLAGS=		-g -Wall -Wno-unused-function -O2
 WRAP_MALLOC=-DUSE_MALLOC_WRAPPERS
+SHOW_PERF=  -DSHOW_PERF
 AR=			ar
-DFLAGS=		-DHAVE_PTHREAD $(WRAP_MALLOC)
+DFLAGS=		-DHAVE_PTHREAD $(WRAP_MALLOC) $(SHOW_PERF)
 LOBJS=		utils.o kthread.o kstring.o ksw.o bwt.o bntseq.o bwa.o bwamem.o bwamem_pair.o bwamem_extra.o malloc_wrap.o \
 			QSufSort.o bwt_gen.o rope.o rle.o is.o bwtindex.o
 AOBJS=		bwashm.o bwase.o bwaseqio.o bwtgap.o bwtaln.o bamlite.o \
diff --git a/bwamem.c b/bwamem.c
index 03e2a05..74b5d53 100644
--- a/bwamem.c
+++ b/bwamem.c
@@ -306,8 +306,15 @@ mem_chain_v mem_chain(const mem_opt_t *opt, const bwt_t *bwt, const bntseq_t *bn
 			mem_chain_t tmp, *lower, *upper;
 			mem_seed_t s;
 			int rid, to_add = 0;
+#ifdef SHOW_PERF
+			int64_t tmp_time = realtime_msec();
+#endif
 			s.rbeg = tmp.pos = bwt_sa(bwt, p->x[0] + k); // this is the base coordinate in the forward-reverse reference
-			s.qbeg = p->info>>32;
+#ifdef SHOW_PERF
+			tmp_time = realtime_msec() - tmp_time;
+			__sync_fetch_and_add(&time_bwt_sa, tmp_time);
+#endif
+			s.qbeg = p->info >> 32;
 			s.score= s.len = slen;
 			rid = bns_intv2rid(bns, s.rbeg, s.rbeg + s.len);
 			if (rid < 0) continue; // bridging multiple reference sequences or the forward-reverse boundary; TODO: split the seed; don't discard it!!!
diff --git a/fastmap.c b/fastmap.c
index 3511263..a3fa4ad 100644
--- a/fastmap.c
+++ b/fastmap.c
@@ -41,6 +41,18 @@
 #include "fmt_idx.h"
 KSEQ_DECLARE(gzFile)
 
+// 记录运行时间的变量
+#ifdef SHOW_PERF
+
+int64_t time_ksw_extend2 = 0,
+		time_ksw_global2 = 0,
+		time_ksw_align2 = 0,
+		time_bwt_smem1a = 0,
+		time_bwt_occ4 = 0,
+		time_bwt_sa = 0;
+
+#endif
+
 extern unsigned char nst_nt4_table[256];
 
 void *kopen(const char *fn, int *_fd);
@@ -406,6 +418,13 @@ int main_mem(int argc, char *argv[])
 		kseq_destroy(aux.ks2);
 		err_gzclose(fp2); kclose(ko2);
 	}
+
+#ifdef SHOW_PERF
+	fprintf(stderr, "\n");
+	fprintf(stderr, "time_bwt_sa: %f s\n", time_bwt_sa / 1000.0 / opt->n_threads);
+	fprintf(stderr, "\n");
+#endif
+
 	return 0;
 }
 
diff --git a/fmt_idx.c b/fmt_idx.c
index f35e3ab..52fe021 100644
--- a/fmt_idx.c
+++ b/fmt_idx.c
@@ -23,5 +23,5 @@ void BuildBwtdFromBwt(bwt_t *bwt, bwtd_t **bwtd_p) {
 
     printf("kmer size: %ld M\n", kmerSize * sizeof(kmer_range_t) / 1024 / 1024);
 
-    exit(0);
+    // exit(0);
 }
\ No newline at end of file
diff --git a/run.sh b/run.sh
index 5664607..4724618 100755
--- a/run.sh
+++ b/run.sh
@@ -1,7 +1,7 @@
-time ./bwa mem -t 1 -M -R @RG\\tID:normal\\tSM:normal\\tPL:illumina\\tLB:normal\\tPG:bwa \
+time ./bwa mem -t 12 -M -R @RG\\tID:normal\\tSM:normal\\tPL:illumina\\tLB:normal\\tPG:bwa \
     /home/zzh/data/reference/human_g1k_v37_decoy.fasta \
     /home/zzh/data/fastq/nm1.fq \
-    /home/zzh/data/fastq/nm2.fq #-o /dev/null
+    /home/zzh/data/fastq/nm2.fq -o /dev/null
 
 # time ./bwa mem -t 1 -M -R @RG\\tID:normal\\tSM:normal\\tPL:illumina\\tLB:normal\\tPG:bwa \
 #    /public/home/zzh/data/reference/human_g1k_v37_decoy.fasta \
diff --git a/utils.c b/utils.c
index 9ceb1be..46181db 100644
--- a/utils.c
+++ b/utils.c
@@ -294,6 +294,13 @@ double realtime(void)
 	return tp.tv_sec + tp.tv_usec * 1e-6;
 }
 
+int64_t realtime_msec(void)
+{
+	struct timeval tv;
+	gettimeofday(&tv, NULL);
+	return (int64_t)1000 * (tv.tv_sec + ((1e-6) * tv.tv_usec));
+}
+
 long peakrss(void)
 {
 	struct rusage r;
diff --git a/utils.h b/utils.h
index d4d4550..1a92542 100644
--- a/utils.h
+++ b/utils.h
@@ -31,6 +31,17 @@
 #include <stdio.h>
 #include <zlib.h>
 
+#ifdef SHOW_PERF
+
+extern int64_t time_ksw_extend2,
+		time_ksw_global2,
+		time_ksw_align2,
+		time_bwt_smem1a,
+		time_bwt_occ4,
+		time_bwt_sa;
+
+#endif
+
 #ifdef __GNUC__
 // Tell GCC to validate printf format string and args
 #define ATTRIBUTE(list) __attribute__ (list)
@@ -86,6 +97,7 @@ extern "C" {
 
 	double cputime(void);
 	double realtime(void);
+	int64_t realtime_msec(void);
 	long peakrss(void);
 
 	void ks_introsort_64 (size_t n, uint64_t *a);