diff --git a/app/vmselect/promql/eval.go b/app/vmselect/promql/eval.go
index e2768124a..84bbd894e 100644
--- a/app/vmselect/promql/eval.go
+++ b/app/vmselect/promql/eval.go
@@ -284,14 +284,45 @@ func evalExpr(ec *EvalConfig, e metricsql.Expr) ([]*timeseries, error) {
 		return rv, nil
 	}
 	if be, ok := e.(*metricsql.BinaryOpExpr); ok {
-		left, err := evalExpr(ec, be.Left)
-		if err != nil {
-			return nil, err
-		}
-		right, err := evalExpr(ec, be.Right)
-		if err != nil {
-			return nil, err
+		// Execute left and right sides of the binary operation in parallel.
+		// This should reduce execution times for heavy queries.
+		// On the other side this can increase CPU and RAM usage when executing heavy queries.
+		// TODO: think on how to limit CPU and RAM usage while leaving short execution times.
+		var left, right []*timeseries
+		var mu sync.Mutex
+		var wg sync.WaitGroup
+		var errGlobal error
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			tss, err := evalExpr(ec, be.Left)
+			mu.Lock()
+			if err != nil {
+				if errGlobal == nil {
+					errGlobal = err
+				}
+			}
+			left = tss
+			mu.Unlock()
+		}()
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			tss, err := evalExpr(ec, be.Right)
+			mu.Lock()
+			if err != nil {
+				if errGlobal == nil {
+					errGlobal = err
+				}
+			}
+			right = tss
+			mu.Unlock()
+		}()
+		wg.Wait()
+		if errGlobal != nil {
+			return nil, errGlobal
 		}
+
 		bf := getBinaryOpFunc(be.Op)
 		if bf == nil {
 			return nil, fmt.Errorf(`unknown binary op %q`, be.Op)
diff --git a/docs/CHANGELOG.md b/docs/CHANGELOG.md
index 4accc62d9..0c5d3d662 100644
--- a/docs/CHANGELOG.md
+++ b/docs/CHANGELOG.md
@@ -2,7 +2,8 @@
 
 # tip
 
-* FEATURE: optimize searching for time series by label filters where individual filters match big number of time series (more than a million). For example, the query `up{job="foobar"}` should work faster if `{job="foobar"}` matches a million of time series, while `up{job="foobar"}` matches much lower number of time series.
+* FEATURE: optimize searching for matching metrics for `metric{<label_filters>}` queries if `<label_filters>` contains at least a single filter. For example, the query `up{job="foobar"}` should find the matching time series much faster than previously.
+* FEATURE: reduce execution times for `q1 <binary_op> q2` queries by executing `q1` and `q2` in parallel.
 * FEATURE: single-node VictoriaMetrics now accepts requests to handlers with `/prometheus` and `/graphite` prefixes such as `/prometheus/api/v1/query`. This improves compatibility with [handlers from VictoriaMetrics cluster](https://victoriametrics.github.io/Cluster-VictoriaMetrics.html#url-format).
 * FEATURE: expose `process_open_fds` and `process_max_fds` metrics. These metrics can be used for alerting when `process_open_fds` reaches `process_max_fds`. See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/402 and https://github.com/VictoriaMetrics/VictoriaMetrics/issues/1037
 * FEATURE: vmalert: add `-datasource.appendTypePrefix` command-line option for querying both Prometheus and Graphite datasource in cluster version of VictoriaMetrics. See [these docs](https://victoriametrics.github.io/vmalert.html#graphite) for details.