diff --git a/AvailableColumnsPanel.c b/AvailableColumnsPanel.c
index b21bd3c3..c0b6af14 100644
--- a/AvailableColumnsPanel.c
+++ b/AvailableColumnsPanel.c
@@ -46,6 +46,11 @@ static HandlerResult AvailableColumnsPanel_eventHandler(Panel* super, int ch) {
          result = HANDLED;
          break;
       }
+      default:
+      {
+         result = Panel_selectByTyping(super, ch);
+         break;
+      }
    }
    return result;
 }
diff --git a/CPUMeter.c b/CPUMeter.c
index 2c4c4dbc..f2a4fe34 100644
--- a/CPUMeter.c
+++ b/CPUMeter.c
@@ -122,38 +122,91 @@ static void CPUMeter_display(Object* cast, RichString* out) {
    }
 }
 
+static void AllCPUsMeter_getRange(Meter* this, int* start, int* count) {
+   int cpus = this->pl->cpuCount;
+   switch(this->type->name[0]) {
+      default:
+      case 'A': // All
+         *start = 0;
+         *count = cpus;
+         break;
+      case 'L': // First Half
+         *start = 0;
+         *count = (cpus+1) / 2;
+         break;
+      case 'R': // Second Half
+         *start = (cpus+1) / 2;
+         *count = cpus / 2;
+         break;
+   }
+}
+
 static void AllCPUsMeter_init(Meter* this) {
    int cpus = this->pl->cpuCount;
    if (!this->drawData)
       this->drawData = calloc(sizeof(Meter*), cpus);
    Meter** meters = (Meter**) this->drawData;
-   for (int i = 0; i < cpus; i++) {
+   int start, count;
+   AllCPUsMeter_getRange(this, &start, &count);
+   for (int i = 0; i < count; i++) {
       if (!meters[i])
-         meters[i] = Meter_new(this->pl, i+1, &CPUMeter);
+         meters[i] = Meter_new(this->pl, start+i+1, &CPUMeter);
       meters[i]->type->init(meters[i]);
    }
-   this->h = Meter_modes[this->mode]->h * cpus;
+   if (this->mode == 0)
+      this->mode = BAR_METERMODE;
+   int h = Meter_modes[this->mode]->h;
+   if (strchr(this->type->name, '2'))
+      this->h = h * ((count+1) / 2);
+   else
+      this->h = h * count;
 }
 
 static void AllCPUsMeter_done(Meter* this) {
-   int cpus = this->pl->cpuCount;
    Meter** meters = (Meter**) this->drawData;
-   for (int i = 0; i < cpus; i++)
+   int start, count;
+   AllCPUsMeter_getRange(this, &start, &count);
+   for (int i = 0; i < count; i++)
       Meter_delete((Object*)meters[i]);
 }
 
 static void AllCPUsMeter_setMode(Meter* this, int mode) {
+   Meter** meters = (Meter**) this->drawData;
    this->mode = mode;
-   int cpus = this->pl->cpuCount;
-   int h = Meter_modes[this->mode]->h;
-   this->h = h * cpus;
+   int h = Meter_modes[mode]->h;
+   int start, count;
+   AllCPUsMeter_getRange(this, &start, &count);
+   for (int i = 0; i < count; i++) {
+      Meter_setMode(meters[i], mode);
+   }
+   if (strchr(this->type->name, '2'))
+      this->h = h * ((count+1) / 2);
+   else
+      this->h = h * count;
 }
 
-static void AllCPUsMeter_draw(Meter* this, int x, int y, int w) {
-   int cpus = this->pl->cpuCount;
+static void DualColCPUsMeter_draw(Meter* this, int x, int y, int w) {
    Meter** meters = (Meter**) this->drawData;
-   for (int i = 0; i < cpus; i++) {
-      Meter_setMode(meters[i], this->mode);
+   int start, count;
+   AllCPUsMeter_getRange(this, &start, &count);
+   int height = (count+1)/2;
+   int startY = y;
+   for (int i = 0; i < height; i++) {
+      meters[i]->draw(meters[i], x, y, (w-2)/2);
+      y += meters[i]->h;
+   }
+   y = startY;
+   for (int i = height; i < count; i++) {
+      meters[i]->draw(meters[i], x+(w-1)/2+2, y, (w-2)/2);
+      y += meters[i]->h;
+   }
+}
+
+static void SingleColCPUsMeter_draw(Meter* this, int x, int y, int w) {
+   Meter** meters = (Meter**) this->drawData;
+   int start, count;
+   AllCPUsMeter_getRange(this, &start, &count);
+   for (int i = 0; i < count; i++) {
       meters[i]->draw(meters[i], x, y, w);
       y += meters[i]->h;
    }
@@ -178,10 +231,81 @@ MeterType AllCPUsMeter = {
    .total = 100.0,
    .attributes = CPUMeter_attributes, 
    .name = "AllCPUs",
-   .uiName = "All CPUs",
+   .uiName = "CPUs (1/1)",
    .caption = "CPU",
-   .draw = AllCPUsMeter_draw,
+   .draw = SingleColCPUsMeter_draw,
    .init = AllCPUsMeter_init,
    .setMode = AllCPUsMeter_setMode,
    .done = AllCPUsMeter_done
 };
+
+MeterType AllCPUs2Meter = {
+   .mode = 0,
+   .items = 1,
+   .total = 100.0,
+   .attributes = CPUMeter_attributes, 
+   .name = "AllCPUs2",
+   .uiName = "CPUs (1&2/2)",
+   .caption = "CPU",
+   .draw = DualColCPUsMeter_draw,
+   .init = AllCPUsMeter_init,
+   .setMode = AllCPUsMeter_setMode,
+   .done = AllCPUsMeter_done
+};
+
+MeterType LeftCPUsMeter = {
+   .mode = 0,
+   .items = 1,
+   .total = 100.0,
+   .attributes = CPUMeter_attributes, 
+   .name = "LeftCPUs",
+   .uiName = "CPUs (1/2)",
+   .caption = "CPU",
+   .draw = SingleColCPUsMeter_draw,
+   .init = AllCPUsMeter_init,
+   .setMode = AllCPUsMeter_setMode,
+   .done = AllCPUsMeter_done
+};
+
+MeterType RightCPUsMeter = {
+   .mode = 0,
+   .items = 1,
+   .total = 100.0,
+   .attributes = CPUMeter_attributes, 
+   .name = "RightCPUs",
+   .uiName = "CPUs (2/2)",
+   .caption = "CPU",
+   .draw = SingleColCPUsMeter_draw,
+   .init = AllCPUsMeter_init,
+   .setMode = AllCPUsMeter_setMode,
+   .done = AllCPUsMeter_done
+};
+
+MeterType LeftCPUs2Meter = {
+   .mode = 0,
+   .items = 1,
+   .total = 100.0,
+   .attributes = CPUMeter_attributes, 
+   .name = "LeftCPUs2",
+   .uiName = "CPUs (1&2/4)",
+   .caption = "CPU",
+   .draw = DualColCPUsMeter_draw,
+   .init = AllCPUsMeter_init,
+   .setMode = AllCPUsMeter_setMode,
+   .done = AllCPUsMeter_done
+};
+
+MeterType RightCPUs2Meter = {
+   .mode = 0,
+   .items = 1,
+   .total = 100.0,
+   .attributes = CPUMeter_attributes, 
+   .name = "RightCPUs2",
+   .uiName = "CPUs (3&4/4)",
+   .caption = "CPU",
+   .draw = DualColCPUsMeter_draw,
+   .init = AllCPUsMeter_init,
+   .setMode = AllCPUsMeter_setMode,
+   .done = AllCPUsMeter_done
+};
+
diff --git a/CPUMeter.h b/CPUMeter.h
index 37f7b431..47ef8882 100644
--- a/CPUMeter.h
+++ b/CPUMeter.h
@@ -34,4 +34,15 @@ extern MeterType CPUMeter;
 
 extern MeterType AllCPUsMeter;
 
+extern MeterType AllCPUs2Meter;
+
+extern MeterType LeftCPUsMeter;
+
+extern MeterType RightCPUsMeter;
+
+extern MeterType LeftCPUs2Meter;
+
+extern MeterType RightCPUs2Meter;
+
+
 #endif
diff --git a/ChangeLog b/ChangeLog
index ae9b57d6..724aaa77 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,7 +1,12 @@
 
-What's new in version 0.9.1
+What's new in version 1.0
 
+* Performance improvements
+* Support for splitting CPU meters into two or four columns
+  (thanks to Wim Heirman)
 * Switch from PLPA, which is now deprecated, to HWLOC.
+* Bring back support for native Linux sched_setaffinity,
+  so we don't have to use HWLOC where we don't need to.
 * Support for typing in user names and column fields in selection panels.
 * Support for UTF-8 tree drawing
   (thanks to Bin Guo)
diff --git a/ColumnsPanel.c b/ColumnsPanel.c
index 94a21232..5c630749 100644
--- a/ColumnsPanel.c
+++ b/ColumnsPanel.c
@@ -60,6 +60,13 @@ static HandlerResult ColumnsPanel_eventHandler(Panel* super, int ch) {
          result = HANDLED;
          break;
       }
+      default:
+      {
+         result = Panel_selectByTyping(super, ch);
+         if (result == BREAK_LOOP)
+            result = IGNORED;
+         break;
+      }
    }
    if (result == HANDLED)
       ColumnsPanel_update(super);
diff --git a/Header.c b/Header.c
index a30c834b..9a14b156 100644
--- a/Header.c
+++ b/Header.c
@@ -124,8 +124,16 @@ MeterModeId Header_readMeterMode(Header* this, int i, HeaderSide side) {
    return meter->mode;
 }
 
-void Header_defaultMeters(Header* this) {
-   Vector_add(this->leftMeters, Meter_new(this->pl, 0, &AllCPUsMeter));
+void Header_defaultMeters(Header* this, int cpuCount) {
+   if (cpuCount > 8) {
+      Vector_add(this->leftMeters, Meter_new(this->pl, 0, &LeftCPUs2Meter));
+      Vector_add(this->rightMeters, Meter_new(this->pl, 0, &RightCPUs2Meter));
+   } else if (cpuCount > 4) {
+      Vector_add(this->leftMeters, Meter_new(this->pl, 0, &LeftCPUsMeter));
+      Vector_add(this->rightMeters, Meter_new(this->pl, 0, &RightCPUsMeter));
+   } else {
+      Vector_add(this->leftMeters, Meter_new(this->pl, 0, &AllCPUsMeter));
+   }
    Vector_add(this->leftMeters, Meter_new(this->pl, 0, &MemoryMeter));
    Vector_add(this->leftMeters, Meter_new(this->pl, 0, &SwapMeter));
    Vector_add(this->rightMeters, Meter_new(this->pl, 0, &TasksMeter));
diff --git a/Header.h b/Header.h
index a74e8579..0fb5a735 100644
--- a/Header.h
+++ b/Header.h
@@ -51,7 +51,7 @@ char* Header_readMeterName(Header* this, int i, HeaderSide side);
 
 MeterModeId Header_readMeterMode(Header* this, int i, HeaderSide side);
 
-void Header_defaultMeters(Header* this);
+void Header_defaultMeters(Header* this, int cpuCount);
 
 void Header_reinit(Header* this);
 
diff --git a/Meter.c b/Meter.c
index 8c3cc50f..97b1dc4e 100644
--- a/Meter.c
+++ b/Meter.c
@@ -130,8 +130,13 @@ MeterType* Meter_types[] = {
    &TasksMeter,
    &UptimeMeter,
    &BatteryMeter,
-   &AllCPUsMeter,
    &HostnameMeter,
+   &AllCPUsMeter,
+   &AllCPUs2Meter,
+   &LeftCPUsMeter,
+   &RightCPUsMeter,
+   &LeftCPUs2Meter,
+   &RightCPUs2Meter,
    NULL
 };
 
@@ -147,9 +152,9 @@ Meter* Meter_new(ProcessList* pl, int param, MeterType* type) {
    this->values = calloc(sizeof(double), type->items);
    this->total = type->total;
    this->caption = strdup(type->caption);
-   Meter_setMode(this, type->mode);
    if (this->type->init)
       this->type->init(this);
+   Meter_setMode(this, type->mode);
    return this;
 }
 
diff --git a/OpenFilesScreen.c b/OpenFilesScreen.c
index 46ecb6d9..b4ec11ae 100644
--- a/OpenFilesScreen.c
+++ b/OpenFilesScreen.c
@@ -151,7 +151,7 @@ static void OpenFilesScreen_scan(OpenFilesScreen* this) {
             free(process->data[i]);
    }
    free(process);
-   Vector_sort(panel->items);
+   Vector_insertionSort(panel->items);
    Panel_setSelected(panel, idx);
 }
 
diff --git a/Panel.c b/Panel.c
index 43432602..56e19488 100644
--- a/Panel.c
+++ b/Panel.c
@@ -93,6 +93,7 @@ void Panel_init(Panel* this, int x, int y, int w, int h, char* type, bool owner)
    this->w = w;
    this->h = h;
    this->eventHandler = NULL;
+   this->eventHandlerBuffer = NULL;
    this->items = Vector_new(type, owner, DEFAULT_SIZE, ListItem_compare);
    this->scrollV = 0;
    this->scrollH = 0;
@@ -108,6 +109,7 @@ void Panel_init(Panel* this, int x, int y, int w, int h, char* type, bool owner)
 
 void Panel_done(Panel* this) {
    assert (this != NULL);
+   free(this->eventHandlerBuffer);
    Vector_delete(this->items);
    RichString_end(this->header);
 }
@@ -405,3 +407,38 @@ bool Panel_onKey(Panel* this, int key) {
    }
    return false;
 }
+
+
+HandlerResult Panel_selectByTyping(Panel* this, int ch) {
+   int size = Panel_size(this);
+   if (!this->eventHandlerBuffer)
+      this->eventHandlerBuffer = calloc(100, 1);
+
+   if (isalnum(ch)) {
+      int len = strlen(this->eventHandlerBuffer);
+      if (len < 99) {
+         this->eventHandlerBuffer[len] = ch;
+         this->eventHandlerBuffer[len+1] = '\0';
+      }
+      for (int try = 0; try < 2; try++) {
+         len = strlen(this->eventHandlerBuffer);
+         for (int i = 0; i < size; i++) {
+            char* cur = ((ListItem*) Panel_get(this, i))->value;
+            while (*cur == ' ') cur++;
+            if (strncasecmp(cur, this->eventHandlerBuffer, len) == 0) {
+               Panel_setSelected(this, i);
+               return HANDLED;
+            }
+         }
+         this->eventHandlerBuffer[0] = ch;
+         this->eventHandlerBuffer[1] = '\0';
+      }
+      return HANDLED;
+   } else if (ch != ERR) {
+      this->eventHandlerBuffer[0] = '\0';
+   }
+   if (ch == 13) {
+      return BREAK_LOOP;
+   }
+   return IGNORED;
+}
diff --git a/Panel.h b/Panel.h
index fb588789..bc838105 100644
--- a/Panel.h
+++ b/Panel.h
@@ -117,4 +117,7 @@ void Panel_draw(Panel* this, bool focus);
 
 bool Panel_onKey(Panel* this, int key);
 
+
+HandlerResult Panel_selectByTyping(Panel* this, int ch);
+
 #endif
diff --git a/Process.c b/Process.c
index 31c87406..02d13966 100644
--- a/Process.c
+++ b/Process.c
@@ -567,6 +567,7 @@ bool Process_setPriority(Process* this, int priority) {
 }
 
 #ifdef HAVE_HWLOC
+
 Affinity* Process_getAffinity(Process* this) {
    hwloc_cpuset_t cpuset = hwloc_bitmap_alloc();
    bool ok = (hwloc_linux_get_tid_cpubind(this->pl->topology, this->pid, cpuset) == 0);
@@ -597,6 +598,31 @@ bool Process_setAffinity(Process* this, Affinity* affinity) {
    hwloc_bitmap_free(cpuset);
    return ok;
 }
+
+#elif HAVE_NATIVE_AFFINITY
+
+Affinity* Process_getAffinity(Process* this) {
+   cpu_set_t cpuset;
+   bool ok = (sched_getaffinity(this->pid, sizeof(cpu_set_t), &cpuset) == 0);
+   if (!ok) return NULL;
+   Affinity* affinity = Affinity_new();
+   for (int i = 0; i < this->pl->cpuCount; i++) {
+      if (CPU_ISSET(i, &cpuset))
+         Affinity_add(affinity, i);
+   }
+   return affinity;
+}
+
+bool Process_setAffinity(Process* this, Affinity* affinity) {
+   cpu_set_t cpuset;
+   CPU_ZERO(&cpuset);
+   for (int i = 0; i < affinity->used; i++) {
+      CPU_SET(affinity->cpus[i], &cpuset);
+   }
+   bool ok = (sched_setaffinity(this->pid, sizeof(unsigned long), &cpuset) == 0);
+   return ok;
+}
+
 #endif
 
 void Process_sendSignal(Process* this, int sgn) {
diff --git a/Process.h b/Process.h
index 403c2aff..c49e0b79 100644
--- a/Process.h
+++ b/Process.h
@@ -197,9 +197,17 @@ void Process_toggleTag(Process* this);
 bool Process_setPriority(Process* this, int priority);
 
 #ifdef HAVE_HWLOC
+
 Affinity* Process_getAffinity(Process* this);
 
 bool Process_setAffinity(Process* this, Affinity* affinity);
+
+#elif HAVE_NATIVE_AFFINITY
+
+Affinity* Process_getAffinity(Process* this);
+
+bool Process_setAffinity(Process* this, Affinity* affinity);
+
 #endif
 
 void Process_sendSignal(Process* this, int sgn);
diff --git a/ProcessList.c b/ProcessList.c
index b681fede..0ba17980 100644
--- a/ProcessList.c
+++ b/ProcessList.c
@@ -178,8 +178,7 @@ ProcessList* ProcessList_new(UsersTable* usersTable) {
    ProcessList* this;
    this = calloc(sizeof(ProcessList), 1);
    this->processes = Vector_new(PROCESS_CLASS, true, DEFAULT_SIZE, Process_compare);
-   this->processTable = Hashtable_new(70, false);
-   assert(Hashtable_count(this->processTable) == Vector_count(this->processes));
+   this->processTable = Hashtable_new(140, false);
    this->usersTable = usersTable;
    
    /* tree-view auxiliary buffers */
@@ -329,7 +328,7 @@ static void ProcessList_buildTree(ProcessList* this, pid_t pid, int level, int i
 
 void ProcessList_sort(ProcessList* this) {
    if (!this->treeView) {
-      Vector_sort(this->processes);
+      Vector_insertionSort(this->processes);
    } else {
       // Save settings
       int direction = this->direction;
@@ -337,7 +336,7 @@ void ProcessList_sort(ProcessList* this) {
       // Sort by PID
       this->sortKey = PID;
       this->direction = 1;
-      Vector_sort(this->processes);
+      Vector_quickSort(this->processes);
       // Restore settings
       this->sortKey = sortKey;
       this->direction = direction;
@@ -448,23 +447,34 @@ static void ProcessList_readIoFile(Process* process, const char* dirname, char*
    unsigned long long last_read = process->io_read_bytes;
    unsigned long long last_write = process->io_write_bytes;
    while (fgets(buffer, 255, file)) {
-      if (sscanf(buffer, "rchar: %llu", &process->io_rchar)) continue;
-      if (sscanf(buffer, "wchar: %llu", &process->io_wchar)) continue;
-      if (sscanf(buffer, "syscr: %llu", &process->io_syscr)) continue;
-      if (sscanf(buffer, "syscw: %llu", &process->io_syscw)) continue;
-      if (sscanf(buffer, "read_bytes: %llu", &process->io_read_bytes)) {
-         process->io_rate_read_bps = 
-            ((double)(process->io_read_bytes - last_read))/(((double)(now - process->io_rate_read_time))/1000);
-         process->io_rate_read_time = now;
-         continue;
+      switch (buffer[0]) {
+      case 'r':
+         if (buffer[1] == 'c')
+            sscanf(buffer, "rchar: %llu", &process->io_rchar);
+         else if (sscanf(buffer, "read_bytes: %llu", &process->io_read_bytes)) {
+            process->io_rate_read_bps = 
+               ((double)(process->io_read_bytes - last_read))/(((double)(now - process->io_rate_read_time))/1000);
+            process->io_rate_read_time = now;
+         }
+         break;
+      case 'w':
+         if (buffer[1] == 'c')
+            sscanf(buffer, "wchar: %llu", &process->io_wchar);
+         else if (sscanf(buffer, "write_bytes: %llu", &process->io_write_bytes)) {
+            process->io_rate_write_bps = 
+               ((double)(process->io_write_bytes - last_write))/(((double)(now - process->io_rate_write_time))/1000);
+            process->io_rate_write_time = now;
+         }
+         break;
+      case 's':
+         if (buffer[5] == 'r')
+            sscanf(buffer, "syscr: %llu", &process->io_syscr);
+         else
+            sscanf(buffer, "syscw: %llu", &process->io_syscw);
+         break;
+      case 'c':
+         sscanf(buffer, "cancelled_write_bytes: %llu", &process->io_cancelled_write_bytes);
       }
-      if (sscanf(buffer, "write_bytes: %llu", &process->io_write_bytes)) {
-         process->io_rate_write_bps = 
-            ((double)(process->io_write_bytes - last_write))/(((double)(now - process->io_rate_write_time))/1000);
-         process->io_rate_write_time = now;
-         continue;
-      }
-      sscanf(buffer, "cancelled_write_bytes: %llu", &process->io_cancelled_write_bytes);
    }
    fclose(file);
 }
diff --git a/Settings.c b/Settings.c
index 3daf3c11..816a7ecb 100644
--- a/Settings.c
+++ b/Settings.c
@@ -55,7 +55,7 @@ static void Settings_readMeterModes(Settings* this, char* line, HeaderSide side)
    String_freeArray(ids);
 }
 
-static bool Settings_read(Settings* this, char* fileName) {
+static bool Settings_read(Settings* this, char* fileName, int cpuCount) {
    // TODO: implement File object and make
    // file I/O object-oriented.
    FILE* fd;
@@ -144,7 +144,7 @@ static bool Settings_read(Settings* this, char* fileName) {
    }
    fclose(fd);
    if (!readMeters) {
-      Header_defaultMeters(this->header);
+      Header_defaultMeters(this->header, cpuCount);
    }
    return true;
 }
@@ -208,7 +208,7 @@ bool Settings_write(Settings* this) {
    return true;
 }
 
-Settings* Settings_new(ProcessList* pl, Header* header) {
+Settings* Settings_new(ProcessList* pl, Header* header, int cpuCount) {
    Settings* this = malloc(sizeof(Settings));
    this->pl = pl;
    this->header = header;
@@ -225,15 +225,15 @@ Settings* Settings_new(ProcessList* pl, Header* header) {
    this->colorScheme = 0;
    this->changed = false;
    this->delay = DEFAULT_DELAY;
-   bool ok = Settings_read(this, this->userSettings);
+   bool ok = Settings_read(this, this->userSettings, cpuCount);
    if (!ok) {
       this->changed = true;
       // TODO: how to get SYSCONFDIR correctly through Autoconf?
       char* systemSettings = String_cat(SYSCONFDIR, "/htoprc");
-      ok = Settings_read(this, systemSettings);
+      ok = Settings_read(this, systemSettings, cpuCount);
       free(systemSettings);
       if (!ok) {
-         Header_defaultMeters(this->header);
+         Header_defaultMeters(this->header, cpuCount);
          pl->hideKernelThreads = true;
          pl->highlightMegabytes = true;
          pl->highlightThreads = false;
diff --git a/Settings.h b/Settings.h
index db3a6469..b75364e3 100644
--- a/Settings.h
+++ b/Settings.h
@@ -32,6 +32,6 @@ void Settings_delete(Settings* this);
 
 bool Settings_write(Settings* this);
 
-Settings* Settings_new(ProcessList* pl, Header* header);
+Settings* Settings_new(ProcessList* pl, Header* header, int cpuCount);
 
 #endif
diff --git a/Vector.c b/Vector.c
index e66e7232..e4d5d5c5 100644
--- a/Vector.c
+++ b/Vector.c
@@ -16,6 +16,8 @@ in the source distribution for its full text.
 
 /*{
 
+#define swap(a_,x_,y_) do{ void* tmp_ = a_[x_]; a_[x_] = a_[y_]; a_[y_] = tmp_; }while(0)
+
 #ifndef DEFAULT_SIZE
 #define DEFAULT_SIZE -1
 #endif
@@ -97,18 +99,83 @@ void Vector_prune(Vector* this) {
    this->items = 0;
 }
 
-void Vector_sort(Vector* this) {
+static int comparisons = 0;
+
+static int partition(Object** array, int left, int right, int pivotIndex, Object_Compare compare) {
+   void* pivotValue = array[pivotIndex];
+   swap(array, pivotIndex, right);
+   int storeIndex = left;
+   for (int i = left; i < right; i++) {
+      comparisons++;
+      if (compare(array[i], pivotValue) <= 0) {
+         swap(array, i, storeIndex);
+         storeIndex++;
+      }
+   }
+   swap(array, storeIndex, right);
+   return storeIndex;
+}
+
+static void quickSort(Object** array, int left, int right, Object_Compare compare) {
+   if (left >= right)
+      return;
+   int pivotIndex = (left+right) / 2;
+   int pivotNewIndex = partition(array, left, right, pivotIndex, compare);
+   quickSort(array, left, pivotNewIndex - 1, compare);
+   quickSort(array, pivotNewIndex + 1, right, compare);
+}
+
+// If I were to use only one sorting algorithm for both cases, it would probably be this one:
+/*
+
+static void combSort(Object** array, int left, int right, Object_Compare compare) {
+   int gap = right - left;
+   bool swapped = true;
+   while ((gap > 1) || swapped) {
+      if (gap > 1) {
+         gap = (int)((double)gap / 1.247330950103979);
+      }
+      swapped = false;
+      for (int i = left; gap + i <= right; i++) {
+         comparisons++;
+         if (compare(array[i], array[i+gap]) > 0) {
+            swap(array, i, i+gap);
+            swapped = true;
+         }
+      }
+   }
+}
+
+*/
+
+static void insertionSort(Object** array, int left, int right, Object_Compare compare) {
+   for (int i = left+1; i <= right; i++) {
+      void* t = array[i];
+      int j = i - 1;
+      while (j >= left) {
+         comparisons++;
+         if (compare(array[j], t) <= 0)
+            break;
+         array[j+1] = array[j];
+         j--;
+      }
+      array[j+1] = t;
+   }
+}
+
+void Vector_quickSort(Vector* this) {
    assert(this->compare);
    assert(Vector_isConsistent(this));
    Object_Compare compare = this->compare;
-   /* Insertion sort works best on mostly-sorted arrays. */
-   for (int i = 1; i < this->items; i++) {
-      int j;
-      void* t = this->array[i];
-      for (j = i-1; j >= 0 && compare(this->array[j], t) > 0; j--)
-         this->array[j+1] = this->array[j];
-      this->array[j+1] = t;
-   }
+   quickSort(this->array, 0, this->items - 1, compare);
+   assert(Vector_isConsistent(this));
+}
+
+void Vector_insertionSort(Vector* this) {
+   assert(this->compare);
+   assert(Vector_isConsistent(this));
+   Object_Compare compare = this->compare;
+   insertionSort(this->array, 0, this->items - 1, compare);
    assert(Vector_isConsistent(this));
 }
 
@@ -205,12 +272,20 @@ void Vector_set(Vector* this, int idx, void* data_) {
    assert(Vector_isConsistent(this));
 }
 
+#ifdef DEBUG
+
 inline Object* Vector_get(Vector* this, int idx) {
    assert(idx < this->items);
    assert(Vector_isConsistent(this));
    return this->array[idx];
 }
 
+#else
+
+#define Vector_get(v_, idx_) ((v_)->array[idx_])
+
+#endif
+
 inline int Vector_size(Vector* this) {
    assert(Vector_isConsistent(this));
    return this->items;
diff --git a/Vector.h b/Vector.h
index fc189bfa..20c0036c 100644
--- a/Vector.h
+++ b/Vector.h
@@ -18,6 +18,8 @@ in the source distribution for its full text.
 #include <assert.h>
 
 
+#define swap(a_,x_,y_) do{ void* tmp_ = a_[x_]; a_[x_] = a_[y_]; a_[y_] = tmp_; }while(0)
+
 #ifndef DEFAULT_SIZE
 #define DEFAULT_SIZE -1
 #endif
@@ -45,7 +47,14 @@ int Vector_count(Vector* this);
 
 void Vector_prune(Vector* this);
 
-void Vector_sort(Vector* this);
+// If I were to use only one sorting algorithm for both cases, it would probably be this one:
+/*
+
+*/
+
+void Vector_quickSort(Vector* this);
+
+void Vector_insertionSort(Vector* this);
 
 void Vector_insert(Vector* this, int idx, void* data_);
 
@@ -59,8 +68,16 @@ void Vector_moveDown(Vector* this, int idx);
 
 void Vector_set(Vector* this, int idx, void* data_);
 
+#ifdef DEBUG
+
 extern Object* Vector_get(Vector* this, int idx);
 
+#else
+
+#define Vector_get(v_, idx_) ((v_)->array[idx_])
+
+#endif
+
 extern int Vector_size(Vector* this);
 
 /*
diff --git a/configure.ac b/configure.ac
index b049215b..514c458d 100644
--- a/configure.ac
+++ b/configure.ac
@@ -2,7 +2,7 @@
 # Process this file with autoconf to produce a configure script.
 
 AC_PREREQ(2.65)
-AC_INIT([htop],[0.9.1],[loderunner@users.sourceforge.net])
+AC_INIT([htop],[1.0],[loderunner@users.sourceforge.net])
 
 # The following two lines are required by hwloc scripts
 AC_USE_SYSTEM_EXTENSIONS
@@ -87,9 +87,12 @@ if test "x$enable_taskstats" = xyes; then
     AC_DEFINE(HAVE_TASKSTATS, 1, [Define if taskstats support enabled.])
 fi
 
-AC_ARG_ENABLE(unicode, [AC_HELP_STRING([--enable-unicode], [enable Unicode support])], ,enable_unicode="no")
+AC_ARG_ENABLE(unicode, [AC_HELP_STRING([--enable-unicode], [enable Unicode support])], ,enable_unicode="yes")
 if test "x$enable_unicode" = xyes; then
-   AC_CHECK_LIB([ncursesw], [refresh], [], [missing_libraries="$missing_libraries libncursesw"])
+   AC_CHECK_LIB([ncursesw], [refresh], [], [
+      missing_libraries="$missing_libraries libncursesw"
+      AC_MSG_ERROR([You may want to use --disable-unicode or install libncursesw.])
+   ])
    AC_CHECK_HEADERS([ncursesw/curses.h],[:],
         [AC_CHECK_HEADERS([ncurses/ncurses.h],[:],
           [AC_CHECK_HEADERS([ncurses.h],[:],[missing_headers="$missing_headers $ac_header"])])])
@@ -110,10 +113,28 @@ AC_CHECK_FILE($PROCDIR/stat,,AC_MSG_ERROR(Cannot find /proc/stat. Make sure you
 AC_CHECK_FILE($PROCDIR/meminfo,,AC_MSG_ERROR(Cannot find /proc/meminfo. Make sure you have a Linux-compatible /proc filesystem mounted. See the file README for help.))
 fi
 
+AC_ARG_ENABLE(native_affinity, [AC_HELP_STRING([--enable-native-affinity], [enable native sched_setaffinity and sched_getaffinity for affinity support, disables hwloc])], ,enable_native_affinity="yes")
+if test "x$enable_native_affinity" = xyes; then
+   AC_MSG_CHECKING([for usable sched_setaffinity])
+   AC_RUN_IFELSE([
+     AC_LANG_PROGRAM([[
+       #include <sched.h>
+       #include <errno.h>
+       static cpu_set_t cpuset;
+       ]], [[
+       CPU_ZERO(&cpuset);
+       sched_setaffinity(0, sizeof(cpu_set_t), &cpuset);
+       if (errno == ENOSYS) return 1;
+       ]])],
+     [AC_DEFINE(HAVE_NATIVE_AFFINITY, 1, [Define if native sched_setaffinity and sched_getaffinity are to be used.])
+      AC_MSG_RESULT([yes])],
+     [AC_MSG_RESULT([no])])
+fi
+
 ##### hwloc
 AC_ARG_ENABLE(system-hwloc, [AC_HELP_STRING([--enable-system-hwloc], [use the copy of hwloc from your system and not the one bundled with the htop sources. (hwloc required)])], ,enable_system_hwloc="no")
 enable_xml=no
-AC_ARG_ENABLE(hwloc, [AC_HELP_STRING([--enable-hwloc], [enable hwloc support for CPU affinity])],, enable_hwloc="yes")
+AC_ARG_ENABLE(hwloc, [AC_HELP_STRING([--enable-hwloc], [enable hwloc support for CPU affinity])],, test "x$enable_native_affinity" = xno && enable_hwloc="yes")
 if test "x$enable_hwloc" = xyes
 then
    if test "x$enable_system_hwloc" = xyes
diff --git a/debug.h b/debug.h
index 1dedb708..ec2f883e 100644
--- a/debug.h
+++ b/debug.h
@@ -8,7 +8,7 @@
    #define realloc(x,s) DebugMemory_realloc(x, s, __FILE__, __LINE__, #x)
    #define strdup(x) DebugMemory_strdup(x, __FILE__, __LINE__)
    #define free(x) DebugMemory_free(x, __FILE__, __LINE__)
-   #define debug_done() DebugMemory_report(); _nc_free_and_exit()
+   #define debug_done() DebugMemory_report(); _nc_freeall()
 
 #elif defined(DEBUGLITE)
 
diff --git a/htop.c b/htop.c
index 6ceabb25..3fedd126 100644
--- a/htop.c
+++ b/htop.c
@@ -193,44 +193,11 @@ static bool changePriority(Panel* panel, int delta) {
    return anyTagged;
 }
 
-static HandlerResult pickWithEnter(Panel* panel, int ch) {
-   int size = Panel_size(panel);
- 
-   if (isalnum(ch)) {
-      int len = strlen(panel->eventHandlerBuffer);
-      if (len < 99) {
-         panel->eventHandlerBuffer[len] = ch;
-         panel->eventHandlerBuffer[len+1] = '\0';
-      }
-      for (int try = 0; try < 2; try++) {
-         len = strlen(panel->eventHandlerBuffer);
-         for (int i = 0; i < size; i++) {
-            char* cur = ((ListItem*) Panel_get(panel, i))->value;
-            while (*cur == ' ') cur++;
-            if (strncasecmp(cur, panel->eventHandlerBuffer, len) == 0) {
-               Panel_setSelected(panel, i);
-               return HANDLED;
-            }
-         }
-         panel->eventHandlerBuffer[0] = ch;
-         panel->eventHandlerBuffer[1] = '\0';
-      }
-      return HANDLED;
-   } else if (ch != ERR) {
-      panel->eventHandlerBuffer[0] = '\0';
-   }
-   if (ch == 13) {
-      return BREAK_LOOP;
-   }
-   return IGNORED;
-}
-
 static Object* pickFromVector(Panel* panel, Panel* list, int x, int y, const char** keyLabels, FunctionBar* prevBar, Header* header) {
    const char* fuKeys[] = {"Enter", "Esc", NULL};
    int fuEvents[] = {13, 27};
-   list->eventHandlerBuffer = calloc(100, 1);
    if (!list->eventHandler)
-      Panel_setEventHandler(list, pickWithEnter);
+      Panel_setEventHandler(list, Panel_selectByTyping);
    ScreenManager* scr = ScreenManager_new(0, y, 0, -1, HORIZONTAL, header, false);
    ScreenManager_add(scr, list, FunctionBar_new(keyLabels, fuKeys, fuEvents), x - 1);
    ScreenManager_add(scr, panel, NULL, -1);
@@ -238,8 +205,6 @@ static Object* pickFromVector(Panel* panel, Panel* list, int x, int y, const cha
    int ch;
    ScreenManager_run(scr, &panelFocus, &ch);
    ScreenManager_delete(scr);
-   free(list->eventHandlerBuffer);
-   list->eventHandlerBuffer = NULL;
    Panel_move(panel, 0, y);
    Panel_resize(panel, COLS, LINES-y-1);
    FunctionBar_draw(prevBar, NULL);
@@ -383,7 +348,7 @@ int main(int argc, char** argv) {
    Process_getMaxPid();
    
    Header* header = Header_new(pl);
-   settings = Settings_new(pl, header);
+   settings = Settings_new(pl, header, pl->cpuCount);
    int headerHeight = Header_calculateHeight(header);
 
    // FIXME: move delay code to settings
@@ -750,7 +715,7 @@ int main(int argc, char** argv) {
          Panel* usersPanel = Panel_new(0, 0, 0, 0, LISTITEM_CLASS, true, ListItem_compare);
          Panel_setHeader(usersPanel, "Show processes of:");
          UsersTable_foreach(ut, addUserToVector, usersPanel);
-         Vector_sort(usersPanel->items);
+         Vector_insertionSort(usersPanel->items);
          ListItem* allUsers = ListItem_new("All users", -1);
          Panel_insert(usersPanel, 0, (Object*) allUsers);
          const char* fuFunctions[] = {"Show    ", "Cancel ", NULL};
@@ -808,7 +773,7 @@ int main(int argc, char** argv) {
          refreshTimeout = 0;
          break;
       }
-#ifdef HAVE_HWLOC
+#if (HAVE_HWLOC || HAVE_NATIVE_AFFINITY)
       case 'a':
       {
          if (pl->cpuCount == 1)