Roflin/rebbarb
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fsp-server: stop clamping GET_FILE replies to 1024, which capped streaming

Browse Source 
Streaming measured 509 KB/s, BELOW the menu's 850 KB/s TCP dump, and the
arithmetic says exactly why.

I had this wrong: QUEUE_SIZE 2 is the number of queued GAME reads, not
concurrent network requests.  fsp_read_queued() sends ONE fsp_get_file and
arms a 1 s alarm; fsp_pop_queue() issues the next only once the response
arrives.  Streaming is therefore STRICTLY ONE REQUEST IN FLIGHT, so

    throughput = bytes_per_reply / round_trip

The server clamped replies to FSP_SPACE (1024), giving 1024 B / ~2.01 ms =
509 KB/s -- the measured figure almost exactly.  But the in-game client asks
for pmtu - 40 (1460 at a 1500 MTU) and accepts it; only fsplib has the 1024
limit, and fsplib never asks for more than that.  So honouring the requested
size is safe for both clients, and at 1440 B (the 1500 MTU budget rounded
down to 32) the same round trip gives ~1.41x, about 700 KB/s.

Replies stay a multiple of 32 unless they are the file tail, because the
client computes data_size = MIN(requested, OSRoundDown32B(reply_len)) and a
non-aligned mid-file reply would be silently truncated.

Also instruments the round trip: with one request in flight the observed
interval IS the RTT, so the report now splits it into total and this server's
median service time and says SERVER-bound or console-bound.  Without that,
"it is slow" has no owner.  And the raw socket now binds ETH_P_IP rather than
ETH_P_ALL, so our own outgoing frames are no longer copied to userspace and
parsed in Python on the path we are trying to shorten.

T13 checks a 1460-byte request yields 1440 (32-aligned) while an fsplib-sized
1024 request still yields exactly 1024.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Dennis Brentjes
2026-10-10 16:39:46 +02:00
parent 2172a85f0e
commit 4e412c8b9b
5 changed files with 60 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scripts/fsp-server.py
+39 -4
View File
@@ -64,8 +64,12 @@ class RawTransport:
def __init__(self, iface, port):
self.port = port
# ETH_P_IP, not ETH_P_ALL: with ALL, every frame on the link --
# including our own outgoing ones -- is copied to userspace and
# parsed in Python, which adds latency to the very round trip we are
# trying to shorten.
self.s = socket.socket(socket.AF_PACKET, socket.SOCK_RAW,
socket.htons(ETH_P_ALL))
socket.htons(ETH_P_IP))
self.s.bind((iface, 0))
self.our_mac = self.s.getsockname()[4][:6]
@@ -73,7 +77,7 @@ class RawTransport:
"""Return (payload, reply_ctx) for a UDP datagram aimed at our port."""
while True:
frame = self.s.recv(65535)
if len(frame) < 34 or struct.unpack("!H", frame[12:14])[0] != ETH_P_IP:
if len(frame) < 34:
continue
ip = frame[14:]
ihl = (ip[0] & 0x0F) * 4
@@ -261,6 +265,14 @@ def main():
"which is what made mini-dhcpd work.")
ap.add_argument("--bind", default="0.0.0.0")
ap.add_argument("--port", type=int, default=21)
ap.add_argument("--max-reply", type=int, default=1440,
help="largest GET_FILE payload, 32-ALIGNED (default 1440 "
"= 1500 MTU - 20 IP - 8 UDP - 12 FSP, rounded down "
"to 32). Streaming is strictly ONE request in "
"flight, so throughput is this divided by the round "
"trip: at ~2 ms, 1024 B gives ~509 KB/s and 1440 B "
"gives ~700 KB/s. fsplib never asks for more than "
"its own FSP_SPACE (1024) so the menu is unaffected.")
ap.add_argument("--quiet-reads", action="store_true",
help="only log the rate, not every read")
ap.add_argument("--trace", action="store_true",
@@ -298,6 +310,7 @@ def main():
print(f" {pkt_[:48].hex()}")
uploads = {} # peer -> staged bytes, committed by CC_INSTALL
service = [] # per-request service time, to split the round trip
served = 0
t_first = None
t_report = time.time()
@@ -467,7 +480,18 @@ def main():
if cmd == CC_GET_FILE:
want = struct.unpack("!H", r["extra"][:2])[0] if len(r["extra"]) >= 2 else 1024
want = min(want, FSP_SPACE)
# ⚠️ Do NOT clamp to FSP_SPACE (1024). The in-game client asks
# for pmtu - 40 (1460 at a 1500 MTU) and accepts it; only
# fsplib has the 1024 limit, and fsplib never asks for more than
# that, so honouring the request is safe for both. Clamping here
# is what held streaming to 509 KB/s: one request in flight means
# throughput is payload / RTT, nothing else.
want = min(want, a.max_reply)
# Keep it a multiple of 32 unless it is the file tail: the client
# does data_size = MIN(requested, OSRoundDown32B(reply_len)), so a
# non-aligned mid-file reply is silently truncated.
if want >= 32:
want &= ~31
if fh_path != real:
if fh:
fh.close()
@@ -478,12 +502,14 @@ def main():
reply(ctx, pack(CC_ERR, SERVER_KEY, seq, pos,
f"{e}\0".encode()), "ERR")
continue
t_req = time.time()
fh.seek(pos)
data = fh.read(want)
reply(ctx, pack(CC_GET_FILE, SERVER_KEY, seq, pos, data), "GET_FILE")
served += len(data)
reads += 1
now = time.time()
service.append(now - t_req)
if t_first is None:
t_first = now
print(f"[{peer[0]}] GET_FILE {path!r} streaming started")
@@ -492,9 +518,18 @@ def main():
f"want={want} -> {len(data)} B")
if now - t_report >= 2.0:
el = max(now - t_first, 1e-9)
svc = sorted(service)
med = svc[len(svc) // 2] * 1000 if svc else 0.0
rtt = (el / reads) * 1000 if reads else 0.0
print(f" streaming {served/el/1024:7.0f} KB/s "
f"{served/2**20:8.1f} MiB in {el:6.1f}s "
f"{reads/el:5.0f} reads/s last pos {pos:#x}")
f"{reads/el:5.0f} reads/s {len(data)} B/reply")
# One request in flight, so the observed interval IS the round
# trip. Splitting it shows whether WE are the slow part.
print(f" round trip {rtt:5.2f} ms, of which this server "
f"{med:5.2f} ms (median) -> "
f"{'SERVER-bound' if med > rtt/2 else 'console-bound'}")
service.clear()
t_report = now
continue