Quellcode durchsuchen

Increase window size to 64kB. Line rate downloads on a 100Mbps link,

anyone?
tags/v0.9.3
Michael Brown vor 17 Jahren
Ursprung
Commit
5b00fbade3
1 geänderte Dateien mit 30 neuen und 9 gelöschten Zeilen
  1. 30
    9
      src/include/gpxe/tcp.h

+ 30
- 9
src/include/gpxe/tcp.h Datei anzeigen

@@ -213,18 +213,39 @@ struct tcp_mss_option {
213 213
 /**
214 214
  * Advertised TCP window size
215 215
  *
216
+ * 
216 217
  * Our TCP window is actually limited by the amount of space available
217 218
  * for RX packets in the NIC's RX ring; we tend to populate the rings
218
- * with far fewer descriptors than a typical driver.  Since we have no
219
- * way of knowing how much of this RX ring space will be available for
220
- * received TCP packets (consider, for example, that they may all be
221
- * consumed by a series of unrelated ARP requests between other
222
- * machines on the network), it is actually not even theoretically
223
- * possible for us to specify an accurate window size.  We therefore
224
- * guess an arbitrary number that is empirically as large as possible
225
- * while avoiding retransmissions due to dropped packets.
219
+ * with far fewer descriptors than a typical driver.  This would
220
+ * result in a desperately small window size, which kills WAN download
221
+ * performance; the maximum bandwidth on any link is limited to
222
+ *
223
+ *    max_bandwidth = ( tcp_window / round_trip_time )
224
+ *
225
+ * With a 4kB window, which probably accurately reflects our amount of
226
+ * buffer space, and a WAN RTT of say 200ms, this gives a maximum
227
+ * achievable bandwidth of 20kB/s, which is not acceptable.
228
+ *
229
+ * We therefore aim to process packets as fast as they arrive, and
230
+ * advertise an "infinite" window.  If we don't process packets as
231
+ * fast as they arrive, then we will drop packets and have to incur
232
+ * the retransmission penalty.
233
+ *
234
+ * Since we don't store out-of-order received packets, the
235
+ * retransmission penalty is that the whole window contents must be
236
+ * resent.
237
+ *
238
+ * We choose to compromise on a window size of 64kB (which is the
239
+ * maximum that can be represented without using TCP options).  This
240
+ * gives a maximum bandwidth of 320kB/s at 200ms RTT, which is
241
+ * probably faster than the actual link bandwidth.  It also limits
242
+ * retransmissions to 64kB, which is reasonable.
243
+ *
244
+ * Finally, since the window goes into a 16-bit field and we cannot
245
+ * actually use 65536, we use a window size of (65536-4) to ensure
246
+ * that payloads remain dword-aligned.
226 247
  */
227
-#define TCP_WINDOW_SIZE	4096
248
+#define TCP_WINDOW_SIZE	( 65536 - 4 )
228 249
 
229 250
 /**
230 251
  * Advertised TCP MSS

Laden…
Abbrechen
Speichern