diff --git a/Makefile.in b/Makefile.in
index cd927cfe..b211f889 100644
--- a/Makefile.in
+++ b/Makefile.in
@@ -222,6 +222,7 @@ TESTS_UNIT := \
 	$(srcroot)test/unit/pa.c \
 	$(srcroot)test/unit/pack.c \
 	$(srcroot)test/unit/pages.c \
+	$(srcroot)test/unit/peak.c \
 	$(srcroot)test/unit/ph.c \
 	$(srcroot)test/unit/prng.c \
 	$(srcroot)test/unit/prof_accum.c \
diff --git a/include/jemalloc/internal/peak.h b/include/jemalloc/internal/peak.h
new file mode 100644
index 00000000..59da3e41
--- /dev/null
+++ b/include/jemalloc/internal/peak.h
@@ -0,0 +1,37 @@
+#ifndef JEMALLOC_INTERNAL_PEAK_H
+#define JEMALLOC_INTERNAL_PEAK_H
+
+typedef struct peak_s peak_t;
+struct peak_s {
+	/* The highest recorded peak value, after adjustment (see below). */
+	uint64_t cur_max;
+	/*
+	 * The difference between alloc and dalloc at the last set_zero call;
+	 * this lets us cancel out the appropriate amount of excess.
+	 */
+	uint64_t adjustment;
+};
+
+#define PEAK_INITIALIZER {0, 0}
+
+static inline uint64_t
+peak_max(peak_t *peak) {
+	return peak->cur_max;
+}
+
+static inline void
+peak_update(peak_t *peak, uint64_t alloc, uint64_t dalloc) {
+	int64_t candidate_max = (int64_t)(alloc - dalloc - peak->adjustment);
+	if (candidate_max > (int64_t)peak->cur_max) {
+		peak->cur_max = candidate_max;
+	}
+}
+
+/* Resets the counter to zero; all peaks are now relative to this point. */
+static inline void
+peak_set_zero(peak_t *peak, uint64_t alloc, uint64_t dalloc) {
+	peak->cur_max = 0;
+	peak->adjustment = alloc - dalloc;
+}
+
+#endif /* JEMALLOC_INTERNAL_PEAK_H */
diff --git a/test/unit/peak.c b/test/unit/peak.c
new file mode 100644
index 00000000..11129785
--- /dev/null
+++ b/test/unit/peak.c
@@ -0,0 +1,47 @@
+#include "test/jemalloc_test.h"
+
+#include "jemalloc/internal/peak.h"
+
+TEST_BEGIN(test_peak) {
+	peak_t peak = PEAK_INITIALIZER;
+	expect_u64_eq(0, peak_max(&peak),
+	    "Peak should be zero at initialization");
+	peak_update(&peak, 100, 50);
+	expect_u64_eq(50, peak_max(&peak),
+	    "Missed update");
+	peak_update(&peak, 100, 100);
+	expect_u64_eq(50, peak_max(&peak), "Dallocs shouldn't change peak");
+	peak_update(&peak, 100, 200);
+	expect_u64_eq(50, peak_max(&peak), "Dallocs shouldn't change peak");
+	peak_update(&peak, 200, 200);
+	expect_u64_eq(50, peak_max(&peak), "Haven't reached peak again");
+	peak_update(&peak, 300, 200);
+	expect_u64_eq(100, peak_max(&peak), "Missed an update.");
+	peak_set_zero(&peak, 300, 200);
+	expect_u64_eq(0, peak_max(&peak), "No effect from zeroing");
+	peak_update(&peak, 300, 300);
+	expect_u64_eq(0, peak_max(&peak), "Dalloc shouldn't change peak");
+	peak_update(&peak, 400, 300);
+	expect_u64_eq(0, peak_max(&peak), "Should still be net negative");
+	peak_update(&peak, 500, 300);
+	expect_u64_eq(100, peak_max(&peak), "Missed an update.");
+	/*
+	 * Above, we set to zero while a net allocator; let's try as a
+	 * net-deallocator.
+	 */
+	peak_set_zero(&peak, 600, 700);
+	expect_u64_eq(0, peak_max(&peak), "No effect from zeroing.");
+	peak_update(&peak, 600, 800);
+	expect_u64_eq(0, peak_max(&peak), "Dalloc shouldn't change peak.");
+	peak_update(&peak, 700, 800);
+	expect_u64_eq(0, peak_max(&peak), "Should still be net negative.");
+	peak_update(&peak, 800, 800);
+	expect_u64_eq(100, peak_max(&peak), "Missed an update.");
+}
+TEST_END
+
+int
+main(void) {
+	return test_no_reentrancy(
+	    test_peak);
+}