diff --git a/thirdparty/photon/photonwrapper.h b/thirdparty/photon/photonwrapper.h
index 4c78937493..fae4179e55 100644
--- a/thirdparty/photon/photonwrapper.h
+++ b/thirdparty/photon/photonwrapper.h
@@ -29,12 +29,6 @@ photon::WorkPool* work_pool;
 #else
 #endif
 
-// using namespace photon;
-// typedef WorkPool PhotonWorkPool;
-// typedef photon::WorkPool PhotonWorkPool;
-// typedef WorkPool PhotonWorkPool;
-// typedef PhotonWorkPool1 PhotonWorkPool;
-// PhotonWorkPool* new_photon_work_pool();
 void* new_photon_work_pool(size_t);
 // void delete_photon_work_pool(void*);
 void delete_photon_work_pool();
@@ -44,6 +38,10 @@ void init_photon_work_pool(size_t);
 // void photon_thread_migrate(void*);
 void photon_thread_create_and_migrate_to_work_pool(void* (* f)(void*), void* arg);
 // void photon_thread_create_and_migrate_to_work_pool(void*, void* (* f)(void*), void* arg);
+int photon_join_current_thread_into_workpool();
+void photon_set_log_output_stdout();
+void photon_set_log_output_stderr();
+void photon_set_log_output_null();
 // direct wrappers to photon functions
 int photon_init_default();
 void photon_thread_create(void* (* f)(void*), void* arg);
diff --git a/vlib/builtin/builtin_d_gcboehm.c.v b/vlib/builtin/builtin_d_gcboehm.c.v
index 9cfa780f0f..d3ca326a26 100644
--- a/vlib/builtin/builtin_d_gcboehm.c.v
+++ b/vlib/builtin/builtin_d_gcboehm.c.v
@@ -185,7 +185,8 @@ fn C.GC_register_my_thread(voidptr) int
 fn C.GC_unregister_my_thread() int
 
 // fn C.GC_get_my_stackbottom(voidptr) voidptr
-// fn C.GC_set_stackbottom(voidptr, voidptr)
+fn C.GC_set_stackbottom(voidptr, voidptr)
+
 // fn C.GC_push_all_stacks()
 
 fn C.GC_add_roots(voidptr, voidptr)
diff --git a/vlib/coroutines/coroutines.c.v b/vlib/coroutines/coroutines.c.v
index 76321075d1..0c3435c445 100644
--- a/vlib/coroutines/coroutines.c.v
+++ b/vlib/coroutines/coroutines.c.v
@@ -25,6 +25,10 @@ fn C.set_photon_thread_stack_allocator(fn (voidptr, int) voidptr, fn (voidptr, v
 fn C.new_photon_work_pool(int) voidptr
 fn C.delete_photon_work_pool()
 fn C.init_photon_work_pool(int)
+fn C.photon_set_log_output_stdout()
+fn C.photon_set_log_output_stderr()
+fn C.photon_set_log_output_null()
+fn C.photon_join_current_thread_into_workpool() int
 fn C.photon_thread_create_and_migrate_to_work_pool(f voidptr, arg voidptr)
 fn C.photon_thread_create(f voidptr, arg voidptr)
 fn C.photon_thread_migrate()
@@ -43,25 +47,15 @@ pub fn sleep(duration time.Duration) {
 
 fn alloc(_ voidptr, stack_size int) voidptr {
 	unsafe {
-		$if gcboehm ? {
-			// TODO: do this only once when the worker thread is created.
-			// I'm currently just doing it here for convenience.
-			// The second time `C.GC_register_my_thread` gets called from the
-			// same thread it will just fail (returning non 0), so it't not
-			// really a problem, it's just not ideal.
-			mut sb := C.GC_stack_base{}
-			C.GC_get_stack_base(&sb)
-			C.GC_register_my_thread(&sb)
-		}
-
 		stack_ptr := malloc(stack_size)
 
 		$if gcboehm ? {
-			// TODO: this wont work if a corouroutine gets moved to a different
+			// TODO: this wont work if a coroutine gets moved to a different
 			// thread, so we are using `C.GC_set_sp_corrector` with our own
 			// corrector function which seems to be the best solution for now.
 			// It would probably be more performant if we could hook into photon's context
 			// switching code (currently not possible) or we write our own implementation.
+			// TODO: update - I'm not sure if what I wrote above is correct. test
 			// C.GC_set_stackbottom(0, stack_ptr)
 
 			C.GC_add_roots(stack_ptr, charptr(stack_ptr) + stack_size)
@@ -74,32 +68,48 @@ fn alloc(_ voidptr, stack_size int) voidptr {
 fn dealloc(_ voidptr, stack_ptr voidptr, stack_size int) {
 	unsafe {
 		$if gcboehm ? {
-			// TODO: only do this once when the worker thread is killed (not in each coroutine)
-			// come up with a solution for this and `C.GC_register_my_thread` (see alloc above)
-			// C.GC_unregister_my_thread()
-
 			C.GC_remove_roots(stack_ptr, charptr(stack_ptr) + stack_size)
 		}
 		free(stack_ptr)
 	}
 }
 
-fn init() {
-	// NOTE `sp_corrector` only works for platforms with the stack growing down
-	// MacOs, Win32 and Linux always have stack growing down.
-	// A proper solution is planned (hopefully) for boehm v8.4.0.
-	C.GC_set_sp_corrector(C.sp_corrector)
-	if C.GC_get_sp_corrector() == unsafe { nil } {
-		panic('stack pointer correction unsupported')
+fn init_photon_vcpu() {
+	C.photon_init_default()
+	$if gcboehm ? {
+		mut sb := C.GC_stack_base{}
+		C.GC_get_stack_base(&sb)
+		C.GC_register_my_thread(&sb)
 	}
+	C.photon_join_current_thread_into_workpool()
+	$if gcboehm ? {
+		C.GC_unregister_my_thread()
+	}
+}
+
+fn init() {
+	$if gcboehm ? {
+		// NOTE `sp_corrector` only works for platforms with the stack growing down
+		// MacOs, Win32 and Linux always have stack growing down.
+		// A proper solution is planned (hopefully) for boehm v8.4.0.
+		C.GC_set_sp_corrector(C.sp_corrector)
+		if C.GC_get_sp_corrector() == unsafe { nil } {
+			panic('stack pointer correction unsupported')
+		}
+	}
+	C.photon_set_log_output_null()
 	C.set_photon_thread_stack_allocator(alloc, dealloc)
 	ret := C.photon_init_default()
 
-	// NOTE: instead of using photon's WorkPool, we could start our own worker threads,
-	// then initialize photon in each of them. If we did that we would need to control
-	// the migration of coroutines across threads etc.
 	if util.nr_jobs >= 1 {
-		C.init_photon_work_pool(util.nr_jobs)
+		// automatic
+		// C.init_photon_work_pool(util.nr_jobs)
+		// manual - pass 0 because we will start our own
+		C.init_photon_work_pool(0)
+		// start our own vcpu's manually
+		for _ in 1 .. util.nr_jobs {
+			spawn init_photon_vcpu()
+		}
 	}
 
 	if ret < 0 {