diff --git a/arch/amd64/hw/timer.c b/arch/amd64/hw/timer.c
index ee45a97..b3c348a 100644
--- a/arch/amd64/hw/timer.c
+++ b/arch/amd64/hw/timer.c
@@ -72,14 +72,14 @@ static uint32_t timer_tick(void *arg) {
         break;
     }
 
-#if defined(DEBUG_COUNTERS)
+//#if defined(DEBUG_COUNTERS)
     static uint64_t last_debug_cycle = 0;
     uint64_t delta = (system_time - last_debug_cycle) / 1000000ULL;
     if (delta >= 1000) {
         sched_debug_cycle(delta);
         last_debug_cycle = system_time;
     }
-#endif
+//#endif
 
     struct list_head *iter, *b_iter;
     uintptr_t irq;
diff --git a/config.h.in b/config.h.in
index 8a79997..903ff2a 100644
--- a/config.h.in
+++ b/config.h.in
@@ -1,7 +1,7 @@
 #pragma once
 
 #define AMD64_KERNEL_STACK      65536
-#define AMD64_MAX_SMP           1
+//#define AMD64_MAX_SMP           4
 
 //#define SLAB_TRACE_ALLOC        1
 #undef SLAB_TRACE_ALLOC
diff --git a/etc/Kernel.makefile b/etc/Kernel.makefile
index cb9777b..6861d05 100644
--- a/etc/Kernel.makefile
+++ b/etc/Kernel.makefile
@@ -51,6 +51,10 @@ KERNEL_OBJ=$(O)/arch/amd64/boot/yboot.o \
 		   $(O)/arch/amd64/syscall_s.o \
 		   $(O)/arch/amd64/syscall.o \
 		   $(O)/arch/amd64/binfmt_elf.o \
+		   $(O)/arch/amd64/smp/smp.o \
+		   $(O)/arch/amd64/smp/ap_code.o \
+		   $(O)/arch/amd64/smp/ipi.o \
+		   $(O)/arch/amd64/smp/irq_ipi_s.o \
 		   $(O)/sys/debug.o \
 		   $(O)/sys/ubsan.o \
 		   $(O)/sys/panic.o \
@@ -127,6 +131,7 @@ DIRS+=$(O)/arch/amd64/hw \
 	  $(O)/drivers/usb \
 	  $(O)/drivers/ata \
 	  $(O)/drivers/pci \
+	  $(O)/arch/amd64/smp \
 	  $(O)/fs \
 	  $(O)/fs/ext2 \
 	  $(O)/sys/block \
@@ -144,6 +149,8 @@ include etc/KernelOptions.makefile
 KERNEL_GIT_VERSION=$(shell git describe --always --tags)
 KERNEL_CFLAGS=-Iinclude \
 			  -Iinclude/arch/amd64/acpica \
+			  -DAMD64_SMP=4 \
+			  -DAMD64_MAX_SMP=4 \
 			  -Iboot \
 			  -fshort-wchar \
 			  -I$(O)/include \
@@ -163,7 +170,8 @@ KERNEL_CFLAGS=-Iinclude \
 			  -Wall \
 			  -Wextra \
 			  -Wno-unused \
-			  -O0 \
+			  -O2 \
+			  -funroll-loops \
 			  -ggdb \
 			  -Werror $(KERNEL_DEF)
 KERNEL_LDFLAGS=-nostdlib \
@@ -202,3 +210,17 @@ $(O)/sys/font/default8x16.psfu.o: etc/default8x16.psfu arch/amd64/incbin.S
 		-DINCBIN_START="_psf_start" \
 		-DINCBIN_END="_psf_end" \
 		-o $@ arch/amd64/incbin.S
+
+$(O)/arch/amd64/smp/ap_code.o: $(O)/arch/amd64/hw/ap_code.bin arch/amd64/incbin.S
+	@echo " RES $@"
+	@$(CC) \
+		-c \
+		-D__ASM__ \
+		-DINCBIN_FILE='"$(O)/arch/amd64/hw/ap_code.bin"' \
+		-DINCBIN_START="amd64_ap_code_start" \
+		-DINCBIN_END="amd64_ap_code_end" \
+		-o $@ arch/amd64/incbin.S
+
+$(O)/arch/amd64/hw/ap_code.bin: arch/amd64/hw/ap_code.nasm
+	@echo "NASM $@"
+	@nasm -fbin -o $@ $<
diff --git a/fs/sysfs.c b/fs/sysfs.c
index 4f7616b..d154881 100644
--- a/fs/sysfs.c
+++ b/fs/sysfs.c
@@ -372,6 +372,8 @@ void sysfs_populate(void) {
     sysfs_add_config_endpoint(dir, "modules", SYSFS_MODE_DEFAULT, 512, NULL, mod_list, NULL);
     sysfs_add_config_endpoint(dir, "debug_serial", SYSFS_MODE_DEFAULT, 32, "serial", debug_config_get, debug_config_set);
     sysfs_add_config_endpoint(dir, "debug_display", SYSFS_MODE_DEFAULT, 32, "display", debug_config_get, debug_config_set);
+    extern size_t sched_ncpus;
+    sysfs_add_config_endpoint(dir, "smp", SYSFS_MODE_DEFAULT, 16, &sched_ncpus, sysfs_config_int64_getter, NULL);
 
     sysfs_add_config_endpoint(NULL, "mem", SYSFS_MODE_DEFAULT, 512, NULL, system_mem_getter, NULL);
 
diff --git a/sys/panic.c b/sys/panic.c
index c5625b4..9908bc1 100644
--- a/sys/panic.c
+++ b/sys/panic.c
@@ -13,7 +13,7 @@ void panicf(const char *fmt, ...) {
 
     asm volatile ("cli");
     va_list args;
-    kfatal("--- Panic ---\n");
+    kfatal("--- Panic (cpu%d) ---\n", get_cpu()->processor_id);
 
     if (sched_ready) {
         debugs(DEBUG_FATAL, "\033[41m");
@@ -35,7 +35,6 @@ void panicf(const char *fmt, ...) {
 #if defined(AMD64_SMP)
     // Send PANIC IPIs to all other CPUs
     size_t cpu = get_cpu()->processor_id;
-    kfatal("cpu%u initiates panic sequence\n", cpu);
     for (size_t i = 0; i < smp_ncpus; ++i) {
         if (i != cpu) {
             amd64_ipi_send(i, IPI_VECTOR_PANIC);
diff --git a/sys/sched.c b/sys/sched.c
index 20532fa..9a3d401 100644
--- a/sys/sched.c
+++ b/sys/sched.c
@@ -25,6 +25,7 @@ static struct thread threads_idle[AMD64_MAX_SMP] = {0};
 static size_t queue_sizes[AMD64_MAX_SMP] = {0};
 int sched_ncpus = 1;
 int sched_ready = 0;
+static int clk = 0;
 
 static spin_t sched_lock = 0;
 
@@ -90,7 +91,11 @@ void sched_queue(struct thread *thr) {
         }
     }
     spin_release_irqrestore(&sched_lock, &irq);
-    sched_queue_to(thr, min_queue_index);
+    if (min_queue_size == 0) {
+        sched_queue_to(thr, (clk++) % sched_ncpus);
+    } else {
+        sched_queue_to(thr, min_queue_index);
+    }
 #else
     sched_queue_to(thr, 0);
 #endif
@@ -156,6 +161,26 @@ void sched_unqueue(struct thread *thr, enum thread_state new_state) {
     }
 }
 
+void sched_debug_cycle(uint64_t ms) {
+    uintptr_t irq;
+    spin_lock_irqsave(&sched_lock, &irq);
+
+    for (int cpu = 0; cpu < sched_ncpus; ++cpu) {
+        debugf(DEBUG_DEFAULT, "cpu%d: ", cpu);
+
+        for (struct thread *thr = queue_heads[cpu]; thr; thr = thr->sched_next) {
+            debugf(DEBUG_DEFAULT, "#%d (%s):<%p> ", thr->proc->pid, thr->proc->name, thr);
+            if (thr->sched_next == queue_heads[cpu]) {
+                break;
+            }
+        }
+
+        debugc(DEBUG_DEFAULT, '\n');
+    }
+
+    spin_release_irqrestore(&sched_lock, &irq);
+}
+
 //#if defined(DEBUG_COUNTERS)
 //static void sched_debug_tree(int level, struct thread *thr, int depth) {
 //    for (int i = 0; i < depth; ++i) {