Megatest

          (debug:print-info 2 *default-log-port* "already connected to a server for " dbfname)
           conn) ;; we are already connected to the server

	;; no conn
        (let* ((sdats (tt:get-server-info-sorted ttdat dbfname))
	       (sdat  (if (null? sdats)
			  #f
			  (let ((indx (max (random (- (length sdats) 1)) 0)))
			    (list-ref sdats indx)))))
	   (debug:print-info 2 *default-log-port* "found sdat " sdat)
           (match sdat
	    ((host port start-time server-id pid dbfname2 servinffile)
	     (assert (equal? dbfname dbfname2) "FATAL: read server info from wrong file.")
             (debug:print-info 2 *default-log-port* "no conn - in match servinffile:" servinffile)
	     (let* ((host-port (conc host":"port))
		    (conn (make-tt-conn

		  (debug:print 0 *default-log-port* "WARNING: server for "dbfname" is busy, cmd is "cmd", will try again in "dly" seconds. This is attempt "(- attemptnum 1))
                  (debug:print 0 *default-log-port* errmsg)
		  (thread-sleep! dly)
		  (tt:handler  ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe server-start-proc)))
	       ((loaded)
		(debug:print 0 *default-log-port* "WARNING: server for "dbfname" is loaded, slowing queries.")
		(tt:backoff-incr (tt-conn-host conn)(tt-conn-port conn))

		;; this would be a good place to force reconnection and connect to a different server
		
		result) ;; (tt:handler  ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe))
	       (else
		result)))
	    (else ;; did not receive properly formated result
	     (if (not res) ;; tt:send-receive telling us that communication failed
		 (let* ((host    (tt-conn-host conn))
			(port    (tt-conn-port conn))

		   ;; (tt:handler ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe)
		   )))))
	(begin
	  (thread-sleep! 1) ;; no conn yet set up, give it a rest and try again
	  (tt:handler ttdat cmd run-id params attemptnum readonly-mode dbfname testsuite mtexe server-start-proc)))))

;; gets server info and appends path to server file
;; sorts by age, --oldest-- now newest first
;;
;; returns list of (host port startseconds server-id servinfofile)
;;
(define (tt:get-server-info-sorted ttdat dbfname)
  (let* ((areapath (tt-areapath ttdat))
	 (sfiles   (tt:find-server areapath dbfname))
	 (sdats    (filter car (map tt:server-get-info sfiles))) ;; first element is #f if the file disappeared while being read
	 (sorted   (sort sdats (lambda (a b)
				 (let* ((starta (list-ref a 2))
					(startb (list-ref b 2)))
				   (if (eq? starta startb)
				       (string>? (list-ref a 3)(list-ref b 3)) ;; if servers started at same time look at server-id
				       (> starta startb))))))
	 (count    0))
    (for-each
     (lambda (rec)
       (if (or (> (length sorted) 1)
	       (common:low-noise-print 120 "server info sorted"))
	   (debug:print 2 *default-log-port* "SERVER #"count": "(string-intersperse (map conc sorted) ", ")))
       (set! count (+ count 1)))

;;======================================================================
;; server
;;======================================================================

(define (tt:sync-dbs ttdat)
  #f)

(define *server-start-requests* '())

;; start the listener and start responding to requests
;;
;; NOTE: organise by dbfname, not run-id so we don't need
;;       to pull in more modules
;;
;; This is the routine called in megatest.scm to start a server. NOTE: sequence is different for main.db vs. X.db

				     (same-host (or (not prime-host) ;; i.e. this is the first host
						    (equal? prime-host host)))
				     (keep-srv  (and good-ping same-host)))
				(if keep-srv	
				    (loop (cdr servrs)
					  host
					  (cons servdat result))
				    (let* ((modtime (file-modification-time servinfofile)))
				      ;; if the .servinfo hasn't been touched in five min
				      ;; we can be pretty sure the server is truly dead
				      (if (> (- (current-seconds) modtime) 360)
					  (handle-exceptions
					   exn
					   (debug:print-info 0 *default-log-port*
							     "Error removing server info file: "servinfofile", "
							     (condition->list exn))
					   (delete-file* servinfofile))
					  (loop (cdr servrs) prime-host result))))))
			     (else
			      ;; can't delete it as we don't have a filename. NOTE: Should never get here.
			      (debug:print-info 0 *default-log-port* "ERROR: bad servinfo record \""servdat"\"")
			      (loop (cdr servrs) prime-host result)) ;; drop 
			     )))))
	       (home-host (if (null? good-srvrs)
			      #f
			      (caar good-srvrs))))
	  ;; by here we have a trustworthy list of servers and we have removed the .servinfo file for any unresponsive servers

	   nosyncdbpath
	   (lambda (nsdb)
	     (dbfile:insert-or-update-process nsdb *procinf*)))
	  (debug:print 0 *default-log-port* "Exiting now.")
	  (exit))))))

(define (tt:keep-running ttdat dbfname dbstruct)

  (thread-sleep! 1)
  
  ;; at this point the server is running and responding to calls, we just monitor
  ;; for db calls and exit if there are none.

  ;; if I am not in the first 3 servers, exit
  (let* ((start-time (current-seconds)))
    (let loop ()
      (let* ((servers   (tt:get-server-info-sorted ttdat dbfname))
	     (home-host (if (null? servers)
			    #f
			    (caar servers)))
	     (my-index  (list-index (lambda (x)
				      (equal? (list-ref x 6)
					      (tt-servinf-file ttdat)))
				    servers))
	     (ok         (cond
			  ((not my-index)
			   (debug:print 0 *default-log-port* "WARNING: Apparently I don't exist.")
			   #f) ;; keep trying or give up?
			  ((not *server-run*)
			   (debug:print 0 *default-log-port* "WARNING: received a stop server from client by remote request.")
			   #f)
			  ((null? servers)
			   (debug:print 0 *default-log-port* "WARNING: no servinfo files found, this cannot be.")
			   #f) ;; not ok
			  ((> my-index 3)

	      (let* ((sinfo-file (tt-servinf-file ttdat)))
		;; (debug:print 0 *default-log-port* "INFO: touching "sinfo-file)
		(set! (file-modification-time sinfo-file) (current-seconds))
		((dbr:dbstruct-sync-proc dbstruct) last-update)
		(dbr:dbstruct-last-update-set! dbstruct curr-secs))))
	
	(if (< (- (current-seconds) (tt-last-access ttdat)) (tt-server-timeout-param))
	    ;; process any requests to start a new server due to load on this one
	    (let* ((requests *server-start-requests*))
	      (set! *server-start-requests* '())
	      (if (> (length requests) 0)
		  (debug:print-info 0 *default-log-port* "Processing "(length requests)" server start requests"))
	      (for-each (lambda (proc)
			  (proc)
			  (thread-sleep! 1))
			requests)
	      (thread-sleep! 5)
	      (loop)))))
    (tt:shutdown-server ttdat)
    (debug:print 0 *default-log-port* "INFO: Server timed out, exiting from tt:keep-running.")))


(define (tt:shutdown-server ttdat)

(define *last-server-start* (make-hash-table))

(define (tt:too-recent-server-start dbfname)
  (let* ((last-run-time (hash-table-ref/default *last-server-start* dbfname #f)))
    (and last-run-time
	 (< (- (current-seconds) last-run-time) 5))))

(define *last-server-start-request-time* 0)
    
;; Given an area path,  start a server process    ### NOTE ### > file 2>&1 
;; if the target-host is set 
;; try running on that host
;;   incidental: rotate logs in logs/ dir.
;;
(define  (tt:server-process-run areapath testsuite mtexe run-id #!key (profile-mode "")) ;; areapath is *toppath* for a given testsuite area
  (assert areapath  "FATAL: tt:server-process-run called without areapath defined.")
  (assert testsuite "FATAL: tt:server-process-run called without testsuite defined.")
  (assert mtexe     "FATAL: tt:server-process-run called without mtexe defined.")
  ;; mtest -server - -m testsuite:ext-tests -db 6.db
  (let* ((dbfname  (dbmod:run-id->dbfname run-id)))
    (if (or (< (- (current-seconds) *last-server-start-request-time*) 5) ;; attempted start less than 5 sec ago
	    (tt:too-recent-server-start dbfname))
	#f
	(let* ((load     (get-normalized-cpu-load))
	       (srvrs    (tt:find-server areapath dbfname))
	       (trying   (length srvrs))
	       (nrun     (number-of-processes-running (conc "mtest.*server.*"testsuite".*"dbfname))))
	  (set! *last-server-start-request-time* (current-seconds))
	  (cond
	   ((> load 4.0)
	    (debug:print 0 *default-log-port* "Normalized load "load" on " (get-host-name) " is over the limit of 4.0. Not starting a server. Please reduce the load on "(get-host-name)" by killing some processes")
	    (thread-sleep! 1) ;; I'm not convinced that a delay here is helpful. -mrw-
	    #f)
	   ((> nrun 100)
	    (debug:print 0 *default-log-port* nrun" servers running on " (get-host-name) ", not starting another.")
	    (thread-sleep! 1)
	    #f)
	   ((> trying 3)
	    (debug:print 0 *default-log-port* trying" servers registered in .servinfo dir. not starting another.")