diff --git a/src/backend/backend.h b/src/backend/backend.h
index ef1bcb88..7cd64a08 100644
--- a/src/backend/backend.h
+++ b/src/backend/backend.h
@@ -292,6 +292,14 @@ struct backend_operations {
 	/// Optional
 	int (*buffer_age)(backend_t *backend_data);
 
+	/// Get the render time of the last frame. If the render is still in progress,
+	/// returns false. The time is returned in `ts`. Frames are delimited by the
+	/// present() calls. i.e. after a present() call, last_render_time() should start
+	/// reporting the time of the just presen1ted frame.
+	///
+	/// Optional, if not available, the most conservative estimation will be used.
+	bool (*last_render_time)(backend_t *backend_data, struct timespec *ts);
+
 	/// The maximum number buffer_age might return.
 	int max_buffer_age;
 
diff --git a/src/backend/gl/egl.c b/src/backend/gl/egl.c
index dd1d4037..770bd5d4 100644
--- a/src/backend/gl/egl.c
+++ b/src/backend/gl/egl.c
@@ -372,6 +372,7 @@ struct backend_operations egl_ops = {
     .deinit = egl_deinit,
     .bind_pixmap = egl_bind_pixmap,
     .release_image = gl_release_image,
+    .prepare = gl_prepare,
     .compose = gl_compose,
     .image_op = gl_image_op,
     .set_image_property = gl_set_image_property,
@@ -380,6 +381,7 @@ struct backend_operations egl_ops = {
     .is_image_transparent = default_is_image_transparent,
     .present = egl_present,
     .buffer_age = egl_buffer_age,
+    .last_render_time = gl_last_render_time,
     .create_shadow_context = gl_create_shadow_context,
     .destroy_shadow_context = gl_destroy_shadow_context,
     .render_shadow = backend_render_shadow_from_mask,
diff --git a/src/backend/gl/gl_common.c b/src/backend/gl/gl_common.c
index 9789eb51..b02572df 100644
--- a/src/backend/gl/gl_common.c
+++ b/src/backend/gl/gl_common.c
@@ -22,6 +22,11 @@
 #include "backend/backend_common.h"
 #include "backend/gl/gl_common.h"
 
+void gl_prepare(backend_t *base, const region_t *reg attr_unused) {
+	auto gd = (struct gl_data *)base;
+	glBeginQuery(GL_TIME_ELAPSED, gd->frame_timing[gd->current_frame_timing]);
+}
+
 GLuint gl_create_shader(GLenum shader_type, const char *shader_str) {
 	log_trace("===\n%s\n===", shader_str);
 
@@ -800,6 +805,9 @@ uint64_t gl_get_shader_attributes(backend_t *backend_data attr_unused, void *sha
 }
 
 bool gl_init(struct gl_data *gd, session_t *ps) {
+	glGenQueries(2, gd->frame_timing);
+	gd->current_frame_timing = 0;
+
 	// Initialize GLX data structure
 	glDisable(GL_DEPTH_TEST);
 	glDepthMask(GL_FALSE);
@@ -1154,10 +1162,33 @@ void gl_present(backend_t *base, const region_t *region) {
 	glDeleteBuffers(2, bo);
 	glDeleteVertexArrays(1, &vao);
 
+	glEndQuery(GL_TIME_ELAPSED);
+	gd->current_frame_timing ^= 1;
+
+	gl_check_err();
+
 	free(coord);
 	free(indices);
 }
 
+bool gl_last_render_time(backend_t *base, struct timespec *ts) {
+	auto gd = (struct gl_data *)base;
+	GLint available = 0;
+	glGetQueryObjectiv(gd->frame_timing[gd->current_frame_timing ^ 1],
+	                   GL_QUERY_RESULT_AVAILABLE, &available);
+	if (!available) {
+		return false;
+	}
+
+	GLuint64 time;
+	glGetQueryObjectui64v(gd->frame_timing[gd->current_frame_timing ^ 1],
+	                      GL_QUERY_RESULT, &time);
+	ts->tv_sec = (long)(time / 1000000000);
+	ts->tv_nsec = (long)(time % 1000000000);
+	gl_check_err();
+	return true;
+}
+
 bool gl_image_op(backend_t *base, enum image_operations op, void *image_data,
                  const region_t *reg_op, const region_t *reg_visible attr_unused, void *arg) {
 	struct backend_image *tex = image_data;
diff --git a/src/backend/gl/gl_common.h b/src/backend/gl/gl_common.h
index 4aad9c0d..f7e721c2 100644
--- a/src/backend/gl/gl_common.h
+++ b/src/backend/gl/gl_common.h
@@ -108,6 +108,8 @@ struct gl_data {
 	gl_shadow_shader_t shadow_shader;
 	GLuint back_texture, back_fbo;
 	GLint back_format;
+	GLuint frame_timing[2];
+	int current_frame_timing;
 	GLuint present_prog;
 
 	bool dithered_present;
@@ -129,6 +131,7 @@ typedef struct session session_t;
 #define GL_PROG_MAIN_INIT                                                                \
 	{ .prog = 0, .unifm_opacity = -1, .unifm_invert_color = -1, .unifm_tex = -1, }
 
+void gl_prepare(backend_t *base, const region_t *reg);
 void x_rect_to_coords(int nrects, const rect_t *rects, coord_t image_dst,
                       int extent_height, int texture_height, int root_height,
                       bool y_inverted, GLint *coord, GLuint *indices);
@@ -142,6 +145,7 @@ void gl_destroy_window_shader(backend_t *backend_data, void *shader);
 uint64_t gl_get_shader_attributes(backend_t *backend_data, void *shader);
 bool gl_set_image_property(backend_t *backend_data, enum image_properties prop,
                            void *image_data, void *args);
+bool gl_last_render_time(backend_t *backend_data, struct timespec *time);
 
 /**
  * @brief Render a region with texture data.
diff --git a/src/backend/gl/glx.c b/src/backend/gl/glx.c
index 109bec94..b4ace133 100644
--- a/src/backend/gl/glx.c
+++ b/src/backend/gl/glx.c
@@ -528,6 +528,7 @@ struct backend_operations glx_ops = {
     .deinit = glx_deinit,
     .bind_pixmap = glx_bind_pixmap,
     .release_image = gl_release_image,
+    .prepare = gl_prepare,
     .compose = gl_compose,
     .image_op = gl_image_op,
     .set_image_property = gl_set_image_property,
@@ -536,6 +537,7 @@ struct backend_operations glx_ops = {
     .is_image_transparent = default_is_image_transparent,
     .present = glx_present,
     .buffer_age = glx_buffer_age,
+    .last_render_time = gl_last_render_time,
     .create_shadow_context = gl_create_shadow_context,
     .destroy_shadow_context = gl_destroy_shadow_context,
     .render_shadow = backend_render_shadow_from_mask,
diff --git a/src/common.h b/src/common.h
index 99ccc8e1..1684b6c2 100644
--- a/src/common.h
+++ b/src/common.h
@@ -249,6 +249,7 @@ typedef struct session {
 	uint64_t last_render;
 
 	struct rolling_avg *frame_time;
+	struct rolling_max *render_stats;
 
 	// === Operation related ===
 	/// Flags related to the root window
diff --git a/src/picom.c b/src/picom.c
index b37ea54d..7d5ab2f2 100644
--- a/src/picom.c
+++ b/src/picom.c
@@ -159,6 +159,19 @@ static inline struct managed_win *find_win_all(session_t *ps, const xcb_window_t
 void queue_redraw(session_t *ps) {
 	// If --benchmark is used, redraw is always queued
 	if (!ps->redraw_needed && !ps->o.benchmark) {
+		if (!ps->first_frame && ps->redirected &&
+		    ps->backend_data->ops->last_render_time) {
+			struct timespec render_time;
+			if (ps->backend_data->ops->last_render_time(ps->backend_data,
+			                                            &render_time)) {
+				int render_time_us = (int)(render_time.tv_sec * 1000000 +
+				                           render_time.tv_nsec / 1000);
+				log_info(
+				    "Last render call took: %d us, rolling_max: %d us",
+				    render_time_us, rolling_max_get_max(ps->render_stats));
+				rolling_max_push(ps->render_stats, render_time_us);
+			}
+		}
 		ev_idle_start(ps->loop, &ps->draw_idle);
 	}
 	ps->redraw_needed = true;
@@ -1817,6 +1830,7 @@ static session_t *session_init(int argc, char **argv, Display *dpy,
 	pixman_region32_init(&ps->screen_reg);
 
 	// TODO(yshui) investigate what's the best window size
+	ps->render_stats = rolling_max_new(128);
 	ps->frame_time = rolling_avg_new(16);
 
 	ps->pending_reply_tail = &ps->pending_reply_head;
@@ -2455,6 +2469,7 @@ static void session_destroy(session_t *ps) {
 	x_free_randr_info(ps);
 
 	rolling_avg_destroy(ps->frame_time);
+	rolling_max_destroy(ps->render_stats);
 
 	// Release custom window shaders
 	free(ps->o.window_shader_fg);