diff --git a/src/README.md b/src/README.md
index 56712a2e1d..283f64a27f 100644
--- a/src/README.md
+++ b/src/README.md
@@ -877,14 +877,14 @@ Returns a copy of *array*. *Array* must not be a "sparse array".
 — Function **lib.htons** *n*
 
 Host to network byte order conversion functions for 32 and 16 bit
-integers *n* respectively.
+integers *n* respectively. Unsigned.
 
 — Function **lib.ntohl** *n*
 
 — Function **lib.ntohs** *n*
 
 Network to host byte order conversion functions for 32 and 16 bit
-integers *n* respectively.
+integers *n* respectively. Unsigned.
 
 — Function **lib.parse** *arg*, *config*
 
diff --git a/src/apps/basic/basic_apps.lua b/src/apps/basic/basic_apps.lua
index 94eb669272..e98956331e 100644
--- a/src/apps/basic/basic_apps.lua
+++ b/src/apps/basic/basic_apps.lua
@@ -42,7 +42,7 @@ end
 function Join:push ()
    for _, inport in ipairs(self.input) do
       while not link.empty(inport) do
-         transmit(self.output.out, receive(inport))
+         transmit(self.output.output, receive(inport))
       end
    end
 end
diff --git a/src/apps/intel/intel10g.lua b/src/apps/intel/intel10g.lua
index 1c60171d69..fd3c1d11a7 100644
--- a/src/apps/intel/intel10g.lua
+++ b/src/apps/intel/intel10g.lua
@@ -23,7 +23,18 @@ local timer = require("core.timer")
 local bits, bitset = lib.bits, lib.bitset
 local band, bor, lshift = bit.band, bit.bor, bit.lshift
 
-num_descriptors = 1024
+local num_descriptors = 1024
+function ring_buffer_size (arg)
+   if not arg then return num_descriptors end
+   local ring_size = assert(tonumber(arg), "bad ring size: " .. arg)
+   if ring_size > 32*1024 then
+      error("ring size too large for hardware: " .. ring_size)
+   end
+   if math.log(ring_size)/math.log(2) % 1 ~= 0 then
+      error("ring size is not a power of two: " .. arg)
+   end
+   num_descriptors = assert(tonumber(arg))
+end
 
 -- Defaults for configurable items
 local default = {
diff --git a/src/apps/intel/intel_app.lua b/src/apps/intel/intel_app.lua
index 2c11a66542..5f4b8a63ff 100644
--- a/src/apps/intel/intel_app.lua
+++ b/src/apps/intel/intel_app.lua
@@ -388,14 +388,14 @@ function mq_sq(pcidevA, pcidevB)
    print("Send traffic from a nicA (SF) to nicB (two VFs)")
    print("The packets should arrive evenly split between the VFs")
    config.app(c, 'sink_ms', basic_apps.Sink)
-   config.link(c, 'source_ms.out -> repeater_ms.input')
+   config.link(c, 'source_ms.output -> repeater_ms.input')
    config.link(c, 'repeater_ms.output -> nicAs.input')
    config.link(c, 'nicAs.output -> sink_ms.in1')
    config.link(c, 'nicBm0.output -> sink_ms.in2')
    config.link(c, 'nicBm1.output -> sink_ms.in3')
    engine.configure(c)
-   link.transmit(engine.app_table.source_ms.output.out, packet.from_string(d1))
-   link.transmit(engine.app_table.source_ms.output.out, packet.from_string(d2))
+   link.transmit(engine.app_table.source_ms.output.output, packet.from_string(d1))
+   link.transmit(engine.app_table.source_ms.output.output, packet.from_string(d2))
 end
 
 -- one multiqueue driver with two apps and do switch stuff
@@ -436,13 +436,13 @@ function mq_sw(pcidevA)
    print ("Send a bunch of packets from Am0")
    print ("half of them go to nicAm1 and half go nowhere")
    config.app(c, 'sink_ms', basic_apps.Sink)
-   config.link(c, 'source_ms.out -> repeater_ms.input')
+   config.link(c, 'source_ms.output -> repeater_ms.input')
    config.link(c, 'repeater_ms.output -> nicAm0.input')
    config.link(c, 'nicAm0.output -> sink_ms.in1')
    config.link(c, 'nicAm1.output -> sink_ms.in2')
    engine.configure(c)
-   link.transmit(engine.app_table.source_ms.output.out, packet.from_string(d1))
-   link.transmit(engine.app_table.source_ms.output.out, packet.from_string(d2))
+   link.transmit(engine.app_table.source_ms.output.output, packet.from_string(d1))
+   link.transmit(engine.app_table.source_ms.output.output, packet.from_string(d2))
 end
 
 function manyreconf(pcidevA, pcidevB, n, do_pf)
@@ -488,14 +488,14 @@ function manyreconf(pcidevA, pcidevB, n, do_pf)
          vlan = 100+i,
       })
       config.app(c, 'sink_ms', basic_apps.Sink)
-      config.link(c, 'source_ms.out -> repeater_ms.input')
+      config.link(c, 'source_ms.output -> repeater_ms.input')
       config.link(c, 'repeater_ms.output -> nicAm0.input')
       config.link(c, 'nicAm0.output -> sink_ms.in1')
       config.link(c, 'nicAm1.output -> sink_ms.in2')
       if do_pf then engine.configure(config.new()) end
       engine.configure(c)
-      link.transmit(engine.app_table.source_ms.output.out, packet.from_string(d1))
-      link.transmit(engine.app_table.source_ms.output.out, packet.from_string(d2))
+      link.transmit(engine.app_table.source_ms.output.output, packet.from_string(d1))
+      link.transmit(engine.app_table.source_ms.output.output, packet.from_string(d2))
       engine.main({duration = 0.1, no_report=true})
       cycles = cycles + 1
       redos = redos + engine.app_table.nicAm1.dev.pf.redos
diff --git a/src/apps/intel/loadgen.lua b/src/apps/intel/loadgen.lua
index dba5b6363c..874cf63fa7 100644
--- a/src/apps/intel/loadgen.lua
+++ b/src/apps/intel/loadgen.lua
@@ -31,7 +31,7 @@ function disable_tx_descriptor_writeback (dev)
    -- Disable writeback of transmit descriptors.
    -- That way our transmit descriptors stay fresh and reusable.
    -- Tell hardware write them to this other memory instead.
-   local bytes = intel10g.num_descriptors * ffi.sizeof(intel10g.rxdesc_t)
+   local bytes = intel10g.ring_buffer_size() * ffi.sizeof(intel10g.rxdesc_t)
    local ptr, phy = memory.dma_alloc(bytes)
    dev.r.TDWBAL(phy % 2^32)
    dev.r.TDWBAH(phy / 2^32)
@@ -40,7 +40,7 @@ end
 function zero_descriptors (dev)
    -- Clear unused descriptors
    local b = memory.dma_alloc(4096)
-   for i = 0, intel10g.num_descriptors-1 do
+   for i = 0, intel10g.ring_buffer_size()-1 do
       -- Make each descriptors point to valid DMA memory but be 0 bytes long.
       dev.txdesc[i].address = memory.virtual_to_physical(b)
       dev.txdesc[i].options = bit.lshift(1, 24) -- End of Packet flag
@@ -64,7 +64,7 @@ function LoadGen:pull ()
    if dev.tdt == 0 then return end
    C.full_memory_barrier()
    if tdh == 0 then
-      dev.r.TDT(intel10g.num_descriptors)
+      dev.r.TDT(intel10g.ring_buffer_size())
    else
       dev.r.TDT(tdh - 1)
    end
diff --git a/src/apps/socket/raw.lua b/src/apps/socket/raw.lua
index 30e8c66897..dfbc4d4471 100644
--- a/src/apps/socket/raw.lua
+++ b/src/apps/socket/raw.lua
@@ -12,13 +12,6 @@ local ethernet = require("lib.protocol.ethernet")
 local ffi = require("ffi")
 local C = ffi.C
 
---ljsyscall returns error as a a cdata instead of a string,
---and the standard assert() doesn't use tostring() on it.
-local function assert(v, ...)
-   if not v then error(tostring(... or 'assertion failed'), 2) end
-   return v, ...
-end
-
 local c, t = S.c, S.types.t
 
 RawSocket = {}
diff --git a/src/apps/socket/unix.lua b/src/apps/socket/unix.lua
index 7bb416c479..8cfc072360 100644
--- a/src/apps/socket/unix.lua
+++ b/src/apps/socket/unix.lua
@@ -9,13 +9,6 @@ local link   = require("core.link")
 local packet = require("core.packet")
 local S      = require("syscall")
 
---ljsyscall returns error as a a cdata instead of a string,
---and the standard assert() doesn't use tostring() on it.
-local function assert(v, ...)
-   if not v then error(tostring(... or 'assertion failed'), 2) end
-   return v, ...
-end
-
 UnixSocket = {}
 UnixSocket.__index = UnixSocket
 
diff --git a/src/apps/tap/tap.lua b/src/apps/tap/tap.lua
index c1bd5fdd8c..884393f241 100644
--- a/src/apps/tap/tap.lua
+++ b/src/apps/tap/tap.lua
@@ -40,35 +40,34 @@ function Tap:new (name)
       sock:close()
       error("Error opening /dev/net/tun: " .. tostring(err))
    end
-   return setmetatable({sock = sock, name = name}, {__index = Tap})
+   return setmetatable({sock = sock, name = name, pkt = packet.allocate()},
+                       {__index = Tap})
 end
 
 function Tap:pull ()
    local l = self.output.output
    if l == nil then return end
    for i=1,engine.pull_npackets do
-      local p = packet.allocate()
-      local len, err = S.read(self.sock, p.data, C.PACKET_PAYLOAD_SIZE)
+      local len, err = S.read(self.sock, self.pkt.data, C.PACKET_PAYLOAD_SIZE)
       -- errno == EAGAIN indicates that the read would of blocked as there is no
       -- packet waiting. It is not a failure.
       if not len and err.errno == const.E.AGAIN then
-         packet.free(p)
          return
       end
       if not len then
-         packet.free(p)
          error("Failed read on " .. self.name .. ": " .. tostring(err))
       end
-      p.length = len
-      link.transmit(l, p)
+      self.pkt.length = len
+      link.transmit(l, self.pkt)
       counter.add(self.shm.input_bytes, len)
       counter.add(self.shm.input_packets)
-      if ethernet:is_mcast(p.data) then
+      if ethernet:is_mcast(self.pkt.data) then
          counter.add(self.shm.input_mcast)
       end
-      if ethernet:is_bcast(p.data) then
+      if ethernet:is_bcast(self.pkt.data) then
          counter.add(self.shm.input_bcast)
       end
+      self.pkt = packet.allocate()
    end
 end
 
diff --git a/src/apps/test/match.lua b/src/apps/test/match.lua
index dee1aacff9..7517d0d342 100644
--- a/src/apps/test/match.lua
+++ b/src/apps/test/match.lua
@@ -91,7 +91,7 @@ function selftest()
    config.app(c, "join", basic_apps.Join)
    config.link(c, "src.output -> join.src")
    config.link(c, "garbage.output -> join.garbage")
-   config.link(c, "join.out -> sink.input")
+   config.link(c, "join.output -> sink.input")
    engine.configure(c)
    engine.main({duration=0.0001})
    assert(#engine.app_table.sink:errors() == 0)
diff --git a/src/apps/vlan/vlan.lua b/src/apps/vlan/vlan.lua
index 056056ecec..64434b389c 100644
--- a/src/apps/vlan/vlan.lua
+++ b/src/apps/vlan/vlan.lua
@@ -147,6 +147,29 @@ function VlanMux:transmit(o, pkt)
    end
 end
 
+function test_tag_untag ()
+   local pkt = packet.from_string(lib.hexundump([[
+      02:aa:aa:aa:aa:aa 02:99:99:99:99:99 08 00 45 00
+      00 54 43 58 40 00 40 01 7c 5c c0 a8 0d 28 ac 14
+      01 10 08 00 9c d4 07 c0 00 01 bc fa e3 57 00 00
+      00 00 f3 44 01 00 00 00 00 00 10 11 12 13 14 15
+      16 17 18 19 1a 1b 1c 1d 1e 1f 20 21 22 23 24 25
+      26 27 28 29 2a 2b 2c 2d 2e 2f 30 31 32 33 34 35
+      36 37
+   ]], 82))
+   local payload = pkt.data + o_ethernet_ethertype
+   local vid = 0
+   for i=0,15 do
+      for j=0,255 do
+         local tag = build_tag(vid)
+         push_tag(pkt, tag)
+         assert(cast(uint32_ptr_t, payload)[0] == tag)
+         vid = vid + 1
+      end
+   end
+   assert(vid == 4096)
+   print("Sucessfully tagged/untagged all potential VLAN tags (0-4095)")
+end
 
 function selftest()
    local app = require("core.app")
@@ -164,4 +187,6 @@ function selftest()
 
    print("source sent: " .. link.stats(app.app_table.source.output.output).input_packets)
    print("sink received: " .. link.stats(app.app_table.sink.input.input).output_packets)
+
+   test_tag_untag()
 end
diff --git a/src/core/lib.lua b/src/core/lib.lua
index a86206a257..68f782a3dd 100644
--- a/src/core/lib.lua
+++ b/src/core/lib.lua
@@ -10,6 +10,8 @@ require("core.clib_h")
 local bit = require("bit")
 local band, bor, bnot, lshift, rshift, bswap =
    bit.band, bit.bor, bit.bnot, bit.lshift, bit.rshift, bit.bswap
+local tonumber = tonumber -- Yes, this makes a performance difference.
+local cast = ffi.cast
 
 -- Returns true if x and y are structurally similar (isomorphic).
 function equal (x, y)
@@ -373,7 +375,8 @@ if ffi.abi("be") then
    function htonl(b) return b end
    function htons(b) return b end
 else
-   function htonl(b) return bswap(b) end
+  -- htonl is unsigned, matching the C version and expectations.
+   function htonl(b) return tonumber(cast('uint32_t', bswap(b))) end
    function htons(b) return rshift(bswap(b), 16) end
 end
 ntohl = htonl
@@ -760,7 +763,12 @@ function selftest ()
    assert(hexdump('\x45\x00\xb6\x7d\x00\xFA\x40\x00\x40\x11'):upper()
          :match('^45.00.B6.7D.00.FA.40.00.40.11$'), "wrong hex dump")
    assert(hexundump('4500 B67D 00FA400040 11', 10)
-          =='\x45\x00\xb6\x7d\x00\xFA\x40\x00\x40\x11', "wrong hex undump")
+         =='\x45\x00\xb6\x7d\x00\xFA\x40\x00\x40\x11', "wrong hex undump")
+   print("Testing ntohl")
+   local raw_val = 0xf0d0b0f0
+   assert(ntohl(raw_val) > 0, "ntohl must be unsigned")
+   assert(ntohl(ntohl(raw_val)) == raw_val, 
+      "calling ntohl twice must return the original value")
 
    -- test parse
    print("Testing parse")
diff --git a/src/core/main.lua b/src/core/main.lua
index 5930142313..611f167152 100644
--- a/src/core/main.lua
+++ b/src/core/main.lua
@@ -19,6 +19,13 @@ local S = require("syscall")
 require("lib.lua.strict")
 require("lib.lua.class")
 
+-- ljsyscall returns error as a cdata instead of a string, and the standard
+-- assert doesn't use tostring on it.
+_G.assert = function (v, ...)
+   if v then return v, ... end
+   error(tostring(... or "assertion failed!"))
+end
+
 -- Reserve names that we want to use for global module.
 -- (This way we avoid errors from the 'strict' module.)
 _G.config, _G.engine, _G.memory, _G.link, _G.packet, _G.timer,
@@ -164,7 +171,7 @@ function selftest ()
 end
 
 -- Fork into worker process and supervisor
-local worker_pid = S.fork()
+local worker_pid = assert(S.fork())
 if worker_pid == 0 then
    -- Worker: Use prctl to ensure we are killed (SIGHUP) when our parent quits
    -- and run main.
@@ -179,16 +186,14 @@ else
    while true do
       -- Read signals from signalfd. Only process the first signal because any
       -- signal causes shutdown.
-      local signals, err = S.util.signalfd_read(signalfd)
-      assert(signals, tostring(err))
+      local signals = assert(S.util.signalfd_read(signalfd))
       for i = 1, #signals do
          local exit_status
          if signals[i].chld then
             -- SIGCHILD means worker state changed: retrieve its status using
             -- waitpid and set exit status accordingly.
-            local status, err, worker =
-               S.waitpid(worker_pid, "stopped,continued")
-            assert(status, tostring(err))
+            local status, _, worker =
+               assert(S.waitpid(worker_pid, "stopped,continued"))
             if     worker.WIFEXITED   then exit_status = worker.EXITSTATUS
             elseif worker.WIFSIGNALED then exit_status = 128 + worker.WTERMSIG
             -- WIFSTOPPED and WIFCONTINUED are ignored.
diff --git a/src/core/memory.lua b/src/core/memory.lua
index ecfba556e1..30f8cc49e9 100644
--- a/src/core/memory.lua
+++ b/src/core/memory.lua
@@ -57,8 +57,7 @@ end
 --- ### HugeTLB: Allocate contiguous memory in bulk from Linux
 
 function allocate_hugetlb_chunk ()
-   local fd, err = syscall.open("/proc/sys/vm/nr_hugepages","rdonly")
-   assert(fd, tostring(err))
+   local fd = assert(syscall.open("/proc/sys/vm/nr_hugepages","rdonly"))
    fd:flock("ex")
    for i =1, 3 do
       local page = C.allocate_huge_page(huge_page_size)
diff --git a/src/doc/default.nix b/src/doc/default.nix
index 673d319931..216ab49790 100644
--- a/src/doc/default.nix
+++ b/src/doc/default.nix
@@ -13,8 +13,8 @@ stdenv.mkDerivation rec {
 
   buildInputs = [ ditaa pandoc git
    (texlive.combine {
-      inherit (texlive) scheme-small luatex luatexbase sectsty titlesec cprotect bigfoot titling droid;
-    })
+      inherit (texlive) scheme-small luatex luatexbase sectsty titlesec cprotect bigfoot titling droid collection-luatex;
+   })
   ];
 
   patchPhase = ''
diff --git a/src/lib/hardware/pci.lua b/src/lib/hardware/pci.lua
index de34a4d5d7..3c3fdf0c09 100644
--- a/src/lib/hardware/pci.lua
+++ b/src/lib/hardware/pci.lua
@@ -142,15 +142,12 @@ function map_pci_memory (device, n, lock)
    if lock then
      assert(f:flock("ex, nb"), "failed to lock " .. filepath)
    end
-   local st, err = f:stat()
-   assert(st, tostring(err))
-   local mem, err = f:mmap(nil, st.size, "read, write", "shared", 0)
-   assert(mem, tostring(err))
+   local st = assert(f:stat())
+   local mem = assert(f:mmap(nil, st.size, "read, write", "shared", 0))
    return ffi.cast("uint32_t *", mem), f
 end
 function close_pci_resource (fd, base)
-   local st, err = fd:stat()
-   assert(st, tostring(err))
+   local st = assert(fd:stat())
    S.munmap(base, st.size)
    fd:close()
 end
@@ -159,8 +156,7 @@ end
 --- mastering is enabled.
 function set_bus_master (device, enable)
    root_check()
-   local f,err = S.open(path(device).."/config", "rdwr")
-   assert(f, tostring(err))
+   local f = assert(S.open(path(device).."/config", "rdwr"))
    local fd = f:getfd()
 
    local value = ffi.new("uint16_t[1]")
diff --git a/src/program/firehose/firehose.lua b/src/program/firehose/firehose.lua
index 0354f1a0ec..6ee05d8e6a 100644
--- a/src/program/firehose/firehose.lua
+++ b/src/program/firehose/firehose.lua
@@ -39,13 +39,6 @@ function run (args)
    end
    function opt.r (arg)
       ring_size = tonumber(arg)
-      if type(ring_size) ~= 'number' then fatal("bad ring size: " .. arg) end
-      if ring_size > 32*1024 then
-         fatal("ring size too large for hardware: " .. ring_size)
-      end
-      if math.log(ring_size)/math.log(2) % 1 ~= 0 then
-         fatal("ring size is not a power of two: " .. arg)
-      end
    end
    args = lib.dogetopt(args, opt, "hHet:i:r:", long_opts)
    if #pciaddresses == 0 then
@@ -79,7 +72,7 @@ int firehose_callback_v1(const char *pciaddr, char **packets, void *rxring,
 
       local intel10g = require("apps.intel.intel10g")
       -- Maximum buffers to avoid packet drops
-      intel10g.num_descriptors = ring_size
+      intel10g.ring_buffer_size(ring_size)
       local nic = intel10g.new_sf({pciaddr=pciaddr})
       nic:open()
 
diff --git a/src/program/lisper/lisper.lua b/src/program/lisper/lisper.lua
index 2b0a5b93b3..953ffef0d7 100644
--- a/src/program/lisper/lisper.lua
+++ b/src/program/lisper/lisper.lua
@@ -28,11 +28,6 @@ local ntohl = lib.ntohl
 local getenv = lib.getenv
 local hexdump = lib.hexdump
 
-local function assert(v, ...) --assert overload because
-   if v then return v, ... end
-   error(tostring((...) or "Assertion failed"), 2)
-end
-
 local function parsehex(s)
    return (s:gsub("[0-9a-fA-F][0-9a-fA-F]", function(cc)
      return string.char(tonumber(cc, 16))
diff --git a/src/program/packetblaster/packetblaster.lua b/src/program/packetblaster/packetblaster.lua
index 4b1c11b77f..ec78197c27 100644
--- a/src/program/packetblaster/packetblaster.lua
+++ b/src/program/packetblaster/packetblaster.lua
@@ -4,10 +4,10 @@ module(..., package.seeall)
 
 local engine    = require("core.app")
 local timer     = require("core.timer")
-local intel10g  = require("apps.intel.intel10g")
 local lib       = require("core.lib")
 local pci       = require("lib.hardware.pci")
 local LoadGen   = require("apps.intel.loadgen").LoadGen
+local Intel82599 = require("apps.intel.intel_app").Intel82599
 
 local function is_device_suitable (pcidev, patterns)
    if not pcidev.usable or pcidev.driver ~= 'apps.intel.intel_app' then
@@ -23,40 +23,55 @@ local function is_device_suitable (pcidev, patterns)
    end
 end
 
-function run_loadgen (c, patterns, duration)
-  local nics = 0
-  pci.scan_devices()
-  for _,device in ipairs(pci.devices) do
-    if is_device_suitable(device, patterns) then
-      nics = nics + 1
-      local name = "nic"..nics
-      config.app(c, name, LoadGen, device.pciaddress)
-      config.link(c, "source."..tostring(nics).."->"..name..".input")
-    end
-  end
-  assert(nics > 0, "<PCI> matches no suitable devices.")
-  engine.busywait = true
-  engine.configure(c)
-  local fn = function ()
-                print("Transmissions (last 1 sec):")
-                engine.report_apps()
-             end
-  local t = timer.new("report", fn, 1e9, 'repeating')
-  timer.activate(t)
-  if duration then engine.main({duration=duration})
-  else             engine.main() end
+function run_loadgen (c, patterns, opts)
+   assert(type(opts) == "table")
+   local use_loadgen = opts.loop == nil or opts.loop
+   local nics = 0
+   pci.scan_devices()
+   for _,device in ipairs(pci.devices) do
+      if is_device_suitable(device, patterns) then
+         nics = nics + 1
+         local name = "nic"..nics
+         if use_loadgen then
+            config.app(c, name, LoadGen, device.pciaddress)
+            config.link(c, "source."..tostring(nics).."->"..name..".input")
+         else
+            config.app(c, name, Intel82599, {pciaddr = device.pciaddress})
+            config.link(c, "source."..tostring(nics).."->"..name..".output")
+         end
+      end
+   end
+   assert(nics > 0, "<PCI> matches no suitable devices.")
+   engine.busywait = true
+   engine.configure(c)
+
+   local report = {}
+   if use_loadgen then
+      local fn = function ()
+         print("Transmissions (last 1 sec):")
+         engine.report_apps()
+      end
+      local t = timer.new("report", fn, 1e9, 'repeating')
+      timer.activate(t)
+   else
+      report = {showlinks = true}
+   end
+
+   if opts.duration then engine.main({duration=opts.duration, report=report})
+   else             engine.main() end
 end
+
 local function show_usage(exit_code)
-  print(require("program.packetblaster.README_inc"))
-  main.exit(exit_code)
+   print(require("program.packetblaster.README_inc"))
+   main.exit(exit_code)
 end
 
 function run(args)
-  if #args == 0 then show_usage(1) end
-  local command = string.gsub(table.remove(args, 1), "-", "_")
-  local modname = ("program.packetblaster.%s.%s"):format(command, command)
-  if not lib.have_module(modname) then
-    show_usage(1)
-  end
-  require(modname).run(args)
+   if #args == 0 then show_usage(1) end
+   local command = string.gsub(table.remove(args, 1), "-", "_")
+   local modname = ("program.packetblaster.%s.%s"):format(command, command)
+   if not lib.have_module(modname) then
+      show_usage(1)
+   end
+   require(modname).run(args)
 end
diff --git a/src/program/packetblaster/replay/README b/src/program/packetblaster/replay/README
index 4bc2a3fe40..64d64da1a2 100644
--- a/src/program/packetblaster/replay/README
+++ b/src/program/packetblaster/replay/README
@@ -2,7 +2,9 @@ Usage: packetblaster replay [OPTIONS] <PCAPFILE> <PCI>...
 
   -D DURATION, --duration DURATION
                              Run for DURATION seconds.
-			     Default: unlimited
+                             Default: unlimited
+  --no-loop                  Do not loop through <PCAPFILE>.
+                             Duration is one second if not set.
   -h, --help
                              Print usage information.
 
diff --git a/src/program/packetblaster/replay/replay.lua b/src/program/packetblaster/replay/replay.lua
index 53f7715e3b..150e654152 100644
--- a/src/program/packetblaster/replay/replay.lua
+++ b/src/program/packetblaster/replay/replay.lua
@@ -9,32 +9,45 @@ local PcapReader = require("apps.pcap.pcap").PcapReader
 local lib        = require("core.lib")
 
 local packetblaster = require("program.packetblaster.packetblaster")
-local usage = require("program.packetblaster.replay.README_inc")
 
 local long_opts = {
    duration     = "D",
-   help         = "h"
+   help         = "h",
+   ["no-loop"]  = 0,
 }
 
+local function show_usage (code)
+   print(require("program.packetblaster.replay.README_inc"))
+   main.exit(code)
+end
+
 function run (args)
-   local opt = {}
-   local duration
    local c = config.new()
-   function opt.D (arg) 
-      duration = assert(tonumber(arg), "duration is not a number!")  
+   local handlers = {}
+   local opts = { loop = true }
+   function handlers.D (arg)
+      opts.duration = assert(tonumber(arg), "duration is not a number!")
    end
-   function opt.h (arg)
-      print(usage)
-      main.exit(1)
+   function handlers.h ()
+      show_usage(0)
+   end
+   handlers["no-loop"] = function ()
+      opts.loop = false
    end
 
-   args = lib.dogetopt(args, opt, "hD:", long_opts)
+   args = lib.dogetopt(args, handlers, "hD:", long_opts)
+   if #args < 2 then show_usage(1) end
    local filename = table.remove(args, 1)
    print (string.format("filename=%s", filename))
    config.app(c, "pcap", PcapReader, filename)
-   config.app(c, "loop", basic_apps.Repeater)
    config.app(c, "source", basic_apps.Tee)
-   config.link(c, "pcap.output -> loop.input")
-   config.link(c, "loop.output -> source.input")
-   packetblaster.run_loadgen(c, args, duration)
+   if opts.loop then
+      config.app(c, "loop", basic_apps.Repeater)
+      config.link(c, "pcap.output -> loop.input")
+      config.link(c, "loop.output -> source.input")
+   else
+      config.link(c, "pcap.output -> source.input")
+      if not opts.duration then opts.duration = 1 end
+   end
+   packetblaster.run_loadgen(c, args, opts)
 end
diff --git a/src/program/packetblaster/synth/README b/src/program/packetblaster/synth/README
index 77630edd71..7109fc09b4 100644
--- a/src/program/packetblaster/synth/README
+++ b/src/program/packetblaster/synth/README
@@ -2,17 +2,17 @@ Usage: packetblaster synth [OPTIONS] <PCI>...
 
   -s SOURCE, --src SOURCE
                              Source MAC-Address.
-			     Default: 00:00:00:00:00:00
+                             Default: 00:00:00:00:00:00
   -d DESTINATION, --dst DESTINATION
                              Destination MAC-Address.
-			     Default: 00:00:00:00:00:00
+                             Default: 00:00:00:00:00:00
   -S SIZES, --sizes SIZES
                              A comma separated list of numbers. Send packets of
                              SIZES bytes.
                              Default: 64
   -D DURATION, --duration DURATION
                              Run for DURATION seconds.
-			     Default: unlimited
+                             Default: unlimited
   -h, --help
                              Print usage information.
 
diff --git a/src/program/packetblaster/synth/synth.lua b/src/program/packetblaster/synth/synth.lua
index d657d5b654..e820c8ce49 100644
--- a/src/program/packetblaster/synth/synth.lua
+++ b/src/program/packetblaster/synth/synth.lua
@@ -8,42 +8,45 @@ local Synth     = require("apps.test.synth").Synth
 local lib       = require("core.lib")
 
 local packetblaster = require("program.packetblaster.packetblaster")
-local usage = require("program.packetblaster.synth.README_inc")
 
 local long_opts = {
    duration     = "D",
    help         = "h",
    src          = "s",
    dst          = "d",
-   sizes        = "S"
+   sizes        = "S",
 }
 
+local function show_usage (code)
+   print(require("program.packetblaster.synth.README_inc"))
+   main.exit(code)
+end
+
 function run (args)
-   local opt = {}
-   local duration
    local c = config.new()
-   function opt.D (arg) 
-      duration = assert(tonumber(arg), "duration is not a number!")  
+   local handlers = {}
+   local opts = {}
+   function handlers.D (arg)
+      opts.duration = assert(tonumber(arg), "duration is not a number!")
    end
-   function opt.h (arg)
-      print(usage)
-      main.exit(1)
+   function handlers.h ()
+      show_usage(0)
    end
 
    local source
    local destination
    local sizes
-   function opt.s (arg) source = arg end
-   function opt.d (arg) destination = arg end
-   function opt.S (arg)
-     sizes = {}
-     for size in string.gmatch(arg, "%d+") do
-       sizes[#sizes+1] = tonumber(size)
-     end
+   function handlers.s (arg) source = arg end
+   function handlers.d (arg) destination = arg end
+   function handlers.S (arg)
+      sizes = {}
+      for size in string.gmatch(arg, "%d+") do
+         sizes[#sizes+1] = tonumber(size)
+      end
    end
 
-   args = lib.dogetopt(args, opt, "hD:s:d:S:", long_opts)
+   args = lib.dogetopt(args, handlers, "hD:s:d:S:", long_opts)
    config.app(c, "source", Synth, { sizes = sizes,
-     src = source, dst = destination })
-   packetblaster.run_loadgen(c, args, duration)
+      src = source, dst = destination })
+   packetblaster.run_loadgen(c, args, opts)
 end