diff --git a/cluster/gce/util.sh b/cluster/gce/util.sh
index 56fe5aa2b92..420b52668f0 100755
--- a/cluster/gce/util.sh
+++ b/cluster/gce/util.sh
@@ -745,16 +745,25 @@ function check-cluster() {
       if [[ ${elapsed} -gt ${KUBE_CLUSTER_INITIALIZATION_TIMEOUT} ]]; then
           echo -e "${color_red}Cluster failed to initialize within ${KUBE_CLUSTER_INITIALIZATION_TIMEOUT} seconds.${color_norm}" >&2
           if [[ ${KUBE_TEST_DEBUG-} =~ ^[yY]$ ]]; then
-            local tmp_log="$(mktemp)"
-            local file
-            for file in /var/log/startupscript.log /var/log/kube-apiserver.log; do
-              echo "${MASTER_NAME}:${file} contents:"
-              if gcloud compute copy-files --project "${PROJECT}" \
-                --zone "${ZONE}" "${MASTER_NAME}:${file}" "${tmp_log}"; then
-                cat "${tmp_log}"
-              fi
+            local savedir="${E2E_REPORT_DIR-}"
+            if [[ -z "${savedir}" ]]; then
+              savedir="$(mktemp -t -d k8s-e2e.XXX)"
+            fi
+            echo "Preserving master logs in ${savedir}"
+            local logdir=/var/log
+            local basename
+            for basename in startupscript kube-apiserver; do
+              # TODO(mml): Perhaps revisit how we name logs for preservation and
+              # centralize an implementation.  Options include putting basename
+              # before hostname and including a timestamp.
+              local src="${logdir}/${basename}.log"
+              local dst="${savedir}/${MASTER_NAME}-${basename}.log"
+              echo "Copying ${MASTER_NAME}:${src}"
+              gcloud compute copy-files \
+                --project "${PROJECT}" --zone "${ZONE}" \
+                "${MASTER_NAME}:${src}" "${dst}" \
+                || true
             done
-            rm -f "${tmp_log}"
           fi
           exit 2
       fi