diff --git a/Core/HLE/sceIo.cpp b/Core/HLE/sceIo.cpp
index dbbf582cec..ecba48f4f4 100644
--- a/Core/HLE/sceIo.cpp
+++ b/Core/HLE/sceIo.cpp
@@ -290,6 +290,8 @@ static void __IoFreeFd(int fd, u32 &error) {
 				CoreTiming::UnscheduleEvent(syncNotifyEvent, ((u64)f->waitingSyncThreads[i] << 32) | fd);
 			}
 
+			PROFILE_THIS_SCOPE("io_rw");
+
 			// Discard any pending results.
 			AsyncIOResult managerResult;
 			ioManager.WaitResult(f->handle, managerResult);
@@ -315,6 +317,8 @@ static void __IoFreeFd(int fd, u32 &error) {
 // TODO: We don't do any of that yet.
 // For now, let's at least delay the callback notification.
 static void __IoAsyncNotify(u64 userdata, int cyclesLate) {
+	PROFILE_THIS_SCOPE("io_rw");
+
 	int fd = (int) userdata;
 
 	u32 error;
@@ -368,6 +372,8 @@ static void __IoAsyncNotify(u64 userdata, int cyclesLate) {
 }
 
 static void __IoSyncNotify(u64 userdata, int cyclesLate) {
+	PROFILE_THIS_SCOPE("io_rw");
+
 	SceUID threadID = userdata >> 32;
 	int fd = (int) (userdata & 0xFFFFFFFF);
 
@@ -616,6 +622,8 @@ static u32 sceKernelStderr() {
 }
 
 u64 __IoCompleteAsyncIO(FileNode *f) {
+	PROFILE_THIS_SCOPE("io_rw");
+
 	if (g_Config.iIOTimingMethod == IOTIMING_REALISTIC) {
 		u64 finishTicks = ioManager.ResultFinishTicks(f->handle);
 		if (finishTicks > CoreTiming::GetTicks()) {
@@ -762,7 +770,7 @@ static u32 npdrmRead(FileNode *f, u8 *data, int size) {
 }
 
 static bool __IoRead(int &result, int id, u32 data_addr, int size, int &us) {
-	PROFILE_THIS_SCOPE("ioread");
+	PROFILE_THIS_SCOPE("io_rw");
 	// Low estimate, may be improved later from the ReadFile result.
 	us = size / 100;
 	if (us < 100) {
@@ -895,6 +903,7 @@ static u32 sceIoReadAsync(int id, u32 data_addr, int size) {
 }
 
 static bool __IoWrite(int &result, int id, u32 data_addr, int size, int &us) {
+	PROFILE_THIS_SCOPE("io_rw");
 	// Low estimate, may be improved later from the WriteFile result.
 	us = size / 100;
 	if (us < 100) {
@@ -1085,6 +1094,7 @@ static u32 npdrmLseek(FileNode *f, s32 where, FileMove whence)
 }
 
 static s64 __IoLseekDest(FileNode *f, s64 offset, int whence, FileMove &seek) {
+	PROFILE_THIS_SCOPE("io_rw");
 	seek = FILEMOVE_BEGIN;
 
 	// Let's make sure this isn't incorrect mid-operation.
diff --git a/UI/DevScreens.cpp b/UI/DevScreens.cpp
index 229ecd9e03..6b106fe891 100644
--- a/UI/DevScreens.cpp
+++ b/UI/DevScreens.cpp
@@ -790,19 +790,19 @@ static const uint32_t nice_colors[] = {
 	0xFFFF40,
 
 	0x40FFFF,
-	0xFF40FF,
+	0xFF70FF,
 	0xc0c0c0,
-	0x8040c0,
+	0xb040c0,
 
-	0x990099,
+	0x184099,
 	0xCC3333,
 	0xFF99CC,
 	0x3399CC,
 
-	0x003366,
 	0x990000,
-	0x33FFFF,
+	0x003366,
 	0xF8F8F8,
+	0x33FFFF,
 };
 
 void DrawProfile(UIContext &ui) {
@@ -810,7 +810,7 @@ void DrawProfile(UIContext &ui) {
 	int numCategories = Profiler_GetNumCategories();
 	int historyLength = Profiler_GetHistoryLength();
 
-	float legendWidth = 100.0f;
+	float legendWidth = 80.0f;
 	for (int i = 0; i < numCategories; i++) {
 		const char *name = Profiler_GetCategoryName(i);
 		float w = 0.0f, h = 0.0f;
@@ -819,11 +819,13 @@ void DrawProfile(UIContext &ui) {
 			legendWidth = w;
 		}
 	}
+	legendWidth += 20.0f;
 
-	float legendStartY = ui.GetBounds().centerY();
+	float rowH = 30.0f;
+	float legendHeight = rowH * numCategories;
+	float legendStartY = legendHeight > ui.GetBounds().centerY() ? ui.GetBounds().y2() - legendHeight : ui.GetBounds().centerY();
 	float legendStartX = ui.GetBounds().x2() - std::min(legendWidth, 200.0f);
 
-	float rowH = 30;
 	const uint32_t opacity = 140 << 24;
 
 	for (int i = 0; i < numCategories; i++) {
@@ -835,7 +837,7 @@ void DrawProfile(UIContext &ui) {
 		ui.DrawTextShadow(name, legendStartX + rowH + 2, y, 0xFFFFFFFF, ALIGN_VBASELINE);
 	}
 
-	float graphWidth = ui.GetBounds().x2() - 120;
+	float graphWidth = ui.GetBounds().x2() - legendWidth - 20.0f;
 	float graphHeight = ui.GetBounds().h * 0.8f;
 
 	std::vector<float> history;
@@ -880,12 +882,14 @@ void DrawProfile(UIContext &ui) {
 		if (area)
 			col = opacity | (col & 0xFFFFFF);
 		UI::Drawable color(col);
+		UI::Drawable outline((opacity >> 1) | 0xFFFFFF);
 
 		if (area) {
 			for (int n = 0; n < historyLength; n++) {
 				float val = history[n];
 				float valY1 = ui.GetBounds().y2() - 10 - (val + total[n]) * scale;
 				float valY2 = ui.GetBounds().y2() - 10 - total[n] * scale;
+				ui.FillRect(outline, Bounds(x, valY2, dx, 1.0f));
 				ui.FillRect(color, Bounds(x, valY1, dx, valY2 - valY1));
 				x += dx;
 				total[n] += val;