Retransmission, Fast Retransmission & Spurious Retransmission

 TCP High-Level Summary

TCP guarantees reliable delivery. So if a segment is not acknowledged within a certain time window, TCP will retransmit it. But not all retransmissions mean a network problem—some are normal behavior, some are optimizations, some are false alarms.

Type	Trigger	Actual Loss?	Wireshark Tag
Retransmission	RTO timeout	Maybe	TCP Retransmission
Fast Retransmission	3 duplicate ACKs	Maybe	TCP Fast Retransmission
Spurious Retransmission	Wrong loss assumption	No	TCP Spurious Retransmission

🔥 When each happens (easy examples) 

Retransmission :  TCP resends a segment because it did not receive an ACK before timeout. Triggered by RTO. Likely packet loss.

• Packet is lost due to congestion.
• Timeout expires.  (Client waits → timeout → resend)

Fast Retransmission : TCP resends early after receiving 3 duplicate ACKs, avoiding waiting for timeout. Triggered by loss or reordering.

• Middle packet is lost
• Later packet arrives
• Receiver sends duplicate ACKs

Spurious Retransmission : Sender retransmits even though the original packet was not lost. Caused by delay, reordering, or path change.

• Packet was just delayed
• TCP guessed wrong 

Real life interpretation

• Retransmission = normal recovery
• Fast Retransmission = recovery but faster
• Spurious Retransmission = false alarm caused by delay 

Which one indicates real network issue?

Most serious

• Regular retransmissions (RTO)
• Caused by actual loss

Sometimes serious

• Fast retransmissions
• Caused by congestion or reordering

Usually not serious

• Spurious retransmissions
• Delay but no loss 

Let’s break them down.

1️⃣ Standard Retransmission (RTO Retransmission) 

This is the basic TCP recovery mechanism. If ACK is not received before the retransmission timeout (RTO) expires, the sender resends the segment.

Why it happens

• Normal packet loss on network
• Congestion
• Receiver load
• Buffer overflow
• Routing issues
• VPN, WAN latency, satellite links 

Trigger : Timeout timer (RTO) expires.

• Frame : 39059, Client  -> Server : Seq#1804834 (no ACK)
• Frame : 39060, Client  → Server : Seq#1804834 (again after timeout) 
• Frame : 39061, Client  → Server : Seq#1804834 (again after timeout)  

How Wireshark shows it : [TCP Retransmission] 

Practical meaning : This is TCP doing its job. 

2️⃣ Fast Retransmission 

Meaning : A performance optimization introduced by TCP Reno/NewReno.

Instead of waiting for timeout expiry, TCP retransmits early when network signals packet loss using duplicate ACKs.

Trigger : Receiver sends 3 duplicate ACKs for the same missing packet:

• Frame : 52318, Client  -> Server : TCP Dup ACK 52200#53
• Frame : 52319, Client  → Server : TCP Dup ACK 52200#54
• Frame : 52320, Client  → Server : TCP Dup ACK 52200#55
• Frame : 52321, Server  → Client : TCP First Retransmission

Reasons

• Out-of-order packets
• Packet loss in middle of stream
• Congestion indication 

 

How Wireshark shows it : [TCP Fast Retransmission]

Practical meaning : Network is unstable OR packets arrived out of order, but TCP handled it quickly.

3️⃣ Spurious Retransmission 

Meaning : A retransmission that happens even though the original packet was NOT lost.

So sender thinks packet is lost → retransmits → but original arrives later anyway.

The cause is generally:

• Delay != loss
• Path change or reordering
• VPN tunneling
• Mobile handover
• Very asymmetric paths
• LTE / 4G / 5G scheduling delays 

 

 

• Frame : 39067, Server-> Client: TCP Spurious Retransmission, Seq : 1291224 
• Frame : 39068, Client  → Server : TCP Dup ACK 39065#1
• Frame : 39069, Server-> Client: TCP Spurious Retransmission, Seq : 1291224
• Frame : 39070, Server  → Client : TCP Dup ACK 39065#2

Common Real Examples

• Satellite internet
• Cloud VPN tunnels
• High-latency TLS handshakes
• Cellular roaming

How Wireshark shows it : [TCP Spurious Retransmission] 

Interpretation : This does NOT mean packet loss. It means TCP believed a loss happened due to delay. 

 Quick Definitions : 

Feature / Behavior	TCP Retransmission	TCP Fast Retransmission (First Retransmission)	TCP Spurious Retransmission
Definition	Resending a packet because sender believes it was lost.	Immediate retransmit triggered by 3 duplicate ACKs (Fast Retransmit).	Retransmission that was not needed (original packet was delivered, just delayed).
Trigger Condition	RTO timeout (Retransmission Timeout) expires OR packet loss detected.	Fast retransmit due to 3 duplicate ACKs indicating likely loss.	Sender misjudges loss due to delay/reordering and retransmits unnecessarily.
Cause	Actual packet loss or severe delay.	Packet lost but detected early using duplicate ACKs.	Packet delay, reordering, delayed ACK, or slow receiver; original packet not lost.
Packet Drop?	Yes, usually indicates a real network drop.	Yes, often real loss but detected faster.	No, original packet arrives later.
Wireshark Label	[TCP Retransmission]	[TCP Fast Retransmission]	[TCP Spurious Retransmission]
Latency Impact	High (RTO can be 200–300 ms or more).	Medium (fast recovery reduces latency impact).	Low–Medium (no real loss, but still adds delay).
Detection Mechanism	Timeout-based (RTO).	Duplicate ACK-based (Fast Retransmit).	DSACK, F-RTO, Eifel algorithm detect that retransmission was unnecessary.
Root Problem	Network drops, congestion, queue overflow.	Slight congestion or isolated packet loss.	Delay, reordering, scheduler stall, NIC buffering—not actual loss.
Severity	🔴 High – real packet loss.	🟠 Medium – actual loss but recovered faster.	🟡 Low/Medium – indicates delay, not loss.
Typical Scenarios	WAN links, overloaded NICs, firewall drops.	High-speed links with occasional packet loss.	Wireless networks, ECMP routing, NIC buffer delay, CPU stall.
ACK Behavior	No ACK received → resend.	3 duplicate ACKs → resend faster.	ACK arrives showing original packet was received → retransmit flagged spurious.

 

Deep Dive into ss (Socket Statistics)

SS stands for Socket Statistics. It’s a Linux utility used to inspect socket connections and their detailed TCP/IP kernel-level statistics, much like the older netstat, but faster, more detailed, and more modern. 

It reads live kernel data directly from:

• /proc/net/tcp
• /proc/net/udp
• /proc/net/unix
• /proc/net/sockstat

 and presents it in human-readable, structured form.

How ss Works Internally

• It queries the kernel’s socket tables using Netlink sockets (specifically NETLINK_INET_DIAG and NETLINK_SOCK_DIAG).
• This means it gets real-time data, including per-connection metrics from the TCP stack.
• The kernel collects metrics like RTT, cwnd, retransmissions, and pacing rate — these are then exposed via ss -i or ss -ti. 

 

Use Case	Description
Connection listing	Show all TCP/UDP connections (like netstat -anp)
Performance debugging	Show RTT, cwnd, retransmissions, pacing rate, etc.
Congestion control analysis	Identify which TCP algorithm (CUBIC, BBR, Reno) is in use
Socket buffer tuning	Observe send/receive queue buildup and detect bottlenecks
Detect retransmissions	Identify packet loss, reordering, DSACKs, or retrans timers
TLS offload visibility	Check whether kernel TLS (tcp-ulp-tls) or kTLS is active
Protocol filtering	Filter by protocol, port, address, or state

Example : 

 

This output is not user-space only — it’s kernel-internal TCP socket telemetry made visible to you. Each field is live, pulled from the kernel’s struct tcp_sock and related substructures.

⚡ Key Advantage

Unlike packet captures (e.g., tcpdump), ss doesn’t require sniffing packets or root network privileges. It reads live socket state directly from the kernel, meaning:

• Zero overhead on traffic
• Instant stats per connection
• Works even on encrypted (TLS) connections 

Why ss Matters

Hook: Start with a common problem, like troubleshooting slow network connections or needing a faster alternative to netstat.

What is ss? Define it as a utility to dump socket statistics, a modern replacement for netstat.

Why use it?

• Speed: It's significantly faster than netstat, especially on systems with many active connections, because it retrieves information directly from the kernel's TCP/IP stack data structures (like /proc/net/tcp, /proc/net/udp, etc.) rather than processing raw output. 
• Features: Provides more detailed information on TCP states, socket options, and kernel-level metrics.

Basic Usage: Getting Started

• The Simplest Command:
• ss (Displays all open non-listening sockets, a good starting point).
• The Essentials (The Go-To Command):
• ss -tunapl
• -t: TCP sockets
• -u: UDP sockets
• -n: Numeric ports (prevents DNS lookups, making it fast)
• -a: All sockets (listening and non-listening)
• -p: Show the Process ID (PID) and name (process that opened the socket)
• -l: Only show listening sockets (often used without -a)

 Focusing Your View: Filtering and Specific Stats

• Show Only Listening Sockets:
• ss -lntu (Listen, numeric, TCP/UDP)
• Display Socket Summary (Quick Health Check):
• ss -s (Shows a summary count of different socket states, like established, closed, etc.)
• Filtering by Protocol and State:
• All ESTABLISHED TCP connections: ss -t state established
• All TIME-WAIT TCP connections: ss -t state time-wait (useful for identifying potential resource exhaustion from many closed connections)
• Specific Port: ss -nt sport = :80 (Source port 80)
• Specific IP/Port: ss -nt dst 192.168.1.1:443 (Destination IP and port)

• Advanced Debugging: Unlocking Deeper Information
• Show Timer Information:
• ss -o (Shows the timer state of a connection, like keepalive, on, or off, and when the next event will occur. Crucial for understanding idle connections.)
• Displaying TCP Information (The Good Stuff):
• ss -tnpmi
• -m: Show Memory usage of the socket.
• -i: Show Internal TCP information (congestion control algorithm, RTT, retransmits, window sizes).
• Example Debugging Output:
• Break down the output of the -i flag:
• wscale:7,7: Window scaling factor (client and server)
• rto:204: Retransmission Timeout (in ms)
• rtt:2.043/4.086: Round-Trip Time (average/variance)
• cwnd:10: Congestion Window size (in segments) 

ss vs. netstat: Why the Switch?

• Performance: Reiterate that netstat requires traversing the /proc/net files, which is slow. ss uses a more efficient method via Netlink sockets.
• Data Richness: ss exposes richer kernel-level data related to TCP congestion, memory, and timers directly, making advanced debugging easier.
• Recommendation: Officially recommend using ss over netstat for modern Linux systems. 

Why ss is Better Than netstat 

Feature	netstat	ss
Speed	Slow (reads from /proc)	Fast (Netlink interface)
IPv6 support	Partial	Full
Per-socket TCP info	Limited	Complete (ss -i)
Congestion control info	No	Yes
Queue lengths (Recv-Q / Send-Q)	Yes	Yes
Process mapping (-p)	Yes	Yes
Filtering capabilities	Basic	Advanced (Berkeley Packet Filter syntax)

 Common Commands :

Command	Description
ss -s	Summary of socket usage by state and type
ss -t	List TCP sockets only
ss -u	List UDP sockets only
ss -l	Show listening sockets
ss -n	Don’t resolve hostnames or ports
ss -p	Show process owning each socket
ss -i	Show internal TCP info (RTT, cwnd, ssthresh, etc.)
ss -ti	Show very detailed TCP internal info (the advanced stats we’ve been analyzing)
ss -o state established	Show established sockets with timers
ss -4, ss -6	Show IPv4 or IPv6 sockets only
ss -a	Show all (listening + non-listening) sockets

Summarizes every major TCP metric from ss -ti, the symptom patterns, their likely causes, and what a healthy baseline looks like :

 

 🩺 How to Use This Table

When analyzing ss -ti output:

• Start with rtt, retrans, and cwnd → detect congestion.
• Then check send-Q, delivery_rate, and snd_wnd → app or flow control issues.
• Review rcv_space, rcv_ooopack, and lost → receiver and link quality.
• Confirm pacing_rate and delivery_rate alignment → transmission health.
• Verify congestion control algorithm suits your environment (CUBIC/BBR).

TCP performance triage + tuning guide built for real-time debugging using ss -ti, netstat, or /proc/net/tcp metrics.

 🔍 Quick Reference – What to Watch in ss -ti 

Metric	Normal Range	Problem Indicator
rtt	<50 ms (LAN), <150 ms (WAN)	rising continuously
rttvar	<10 ms	>30 ms means jitter
cwnd	grows steadily	stuck or oscillates
retrans	<0.1%	>1% indicates loss
send-Q	small, fluctuating	growing → app/ACK delay
delivery_rate	near link speed	<50% of link → bottleneck

Example :

Example 2 :

 

 Go parameter by parameter :

Category	Field	Description	Default Unit	Example Value	Use / Functionality
Connection Info	State	TCP connection state	N/A	ESTAB	Shows connection lifecycle (ESTAB, LISTEN, TIME-WAIT)
	Recv-Q	Bytes in receive queue (the count of bytes not yet copied by the user program i,e not yet read by application)	bytes	0	Detects receiver buffer backlog / app slowness
	Send-Q	Bytes in send queue (the count of bytes not yet acknowledged by the remote host i,e not yet ACKed )	bytes	4	Detects sender congestion or buffer saturation
	Local Address:Port	Source IP and port (source of the connection)	N/A	192.168.50.155:49616	Identifies local endpoint
	Peer Address:Port	Destination IP and port(destination of the connection)	N/A	192.168.50.208:11222	Identifies remote endpoint
Congestion Control & Window Scaling	Congestion Control	Algorithm (CUBIC, Reno, BBR)	N/A	cubic	Determines congestion control behavior
	wscale	Send/Receive window scale factors i,e first number (6) is for local (send) side, second (7) is for the remote (receive) side	exponent	7,2	Determines max window for high-BW links
	cwnd	Congestion window. The maximum number of segments the local host is allowed to send before waiting for an ACK.	MSS	7	Indicates congestion; small → throttling
	ssthresh	Slow start threshold. The congestion window size at which the congestion avoidance phase begins.	MSS / bytes	7	Switch between slow start & congestion avoidance
	snd_wnd	Sender’s view of peer advertised window. The maximum amount of unacknowledged data the local host can send (the peer's advertised receive window).	bytes	1,103,872	Flow control; avoid overwhelming receiver
	rcv_ssthresh	Receiver slow start threshold. Used by the local host to limit its advertised receive window when coming out of a slow-start phase.	bytes	66,607	Used in slow start; impacts congestion control
	max_window	Maximum allowed window	bytes	16,777,216	Hard limit for TCP window scaling
RTT, Timers & Pacing	rtt	Smoothed round-trip time (current/variance). First value(1.499) is the Smoothed RTT (SRTT), and the second value(0.962) is the Mean Deviation (RTT variance).	ms	1.499/0.962	Latency measurement; impacts RTO & congestion control
	minrtt	Minimum observed RTT	ms	0.189	Baseline latency; detects jitter
	rto	Retransmission timeout I,e If an ACK isn't received within this time, the segment is retransmitted.	ms	202	Timeout for retransmissions; tuning & loss detection
	ato	Delayed ACK timeout.The time before a TCP stack sends a standalone ACK.	ms	40	Determines when delayed ACKs are sent
	lastsnd	Timestamp of last sent segment. Time since the last data segment was sent.	ms	3444	Debugging stalled connections
	lastrcv	Timestamp of last received segment. Time since the last segment was received.	ms	3409	Debugging stalled connections
	lastack	Timestamp of last ACK received. Time since the last ACK was sent by the local host	ms	3409	Check acknowledgment lag
	busy	Socket active time.Total time the connection has been busy (sending or waiting for an ACK).	ms	801	Measures socket utilization; idle vs active
	pacing_rate	Kernel pacing rate.The rate at which the kernel is scheduled to pace out the packets (rate limiting the output).	bps	65.4 Mbps	Monitor throughput; useful for pacing-based CC
	send	Current send rate. The current sending rate the TCP stack calculates it is achieving.	bps	54.5 Mbps	Real-time send throughput; app-limited vs network-limited
	delivery_rate	Rate of bytes ACKed.The actual observed rate at which the network is delivering data (usually a more accurate measure of throughput than send)	bps	20.4 Mbps	Effective throughput; low value → packet loss
	delivered	Segments successfully delivered. Total number of data segments successfully delivered to the receiver.	count	514	Shows actual data transfer progress
	timer	Retransmit / delayed ACK timer info	ms / struct	N/A	Debugging retransmit scheduling and delayed ACK
MSS / PMTU / Segments	mss	Maximum segment size (payload only).The largest amount of data that a TCP segment can hold.	bytes	1460	Determines segmentization; affects throughput
	pmtu	Path MTU. The largest packet size that can traverse the entire path without fragmentation.	bytes	1500	Detects MTU restrictions to avoid fragmentation
	rcvmss	Peer advertised MSS. The maximum segment size the remote host reported it is willing to receive.	bytes	536	Receiver MSS; affects sender segment sizing
	advmss	Our advertised MSS. The MSS the local host advertised to the peer (usually PMTU minus IP/TCP header size)	bytes	1460	Determines max payload per segment
	segs_out	Total segments sent.Total segments sent (including data, retransmissions, ACKs, etc.).	count	831	Monitor traffic volume
	segs_in	Total segments received	count	589	Monitor inbound traffic
	data_segs_out	Segments carrying payload sent. Total segments sent that contained actual user data.	count	514	Separate data from control segments
	data_segs_in	Segments carrying payload received. Total segments received that contained actual user data.	count	378	Separate data from control segments
Byte Counters	bytes_sent	Total payload bytes sent.Total bytes sent on the connection.	bytes	548,587	Throughput measurement & retransmissions
	bytes_retrans	Bytes retransmitted.Total bytes retransmitted (sent again due to loss).	bytes	3,547	Indicates packet loss and network issues
	bytes_acked	Bytes acknowledged.Total bytes acknowledged by the remote host.	bytes	545,041	Confirms successful delivery
	bytes_received	Bytes received. Total bytes received on the connection.	bytes	4,240	Confirms data reception
Reliability / Loss / Reordering	lost	Segments marked lost	count	4	Detect packet drops & network reliability
	unacked	Segments sent but not yet acknowledged.Number of segments sent but not yet acknowledged.	count	4	In-flight data; too high → congestion
	notsent	Bytes waiting in send buffer.Amount of data queued for sending but not yet sent (often due to congestion control limits).	bytes	66,144	Monitor buffer saturation
	reordering	Number of reordering events.An estimated count of how many segments have been reordered (arrived out of sequence).	count	20	Indicates packet reordering in network
	rcv_ooopack	Out-of-order segments received	count	24	May affect ACK behavior & retransmissions
	dsack_dups	Duplicate segments detected by DSACK.Number of times a Duplicate Selective Acknowledgment (DSACK) was received. Indicates out-of-order arrival or loss.	count	3	Indicates duplicate delivery / network issues
	retrans	Pending / total retransmissions.The first number (0) is the number of segments currently in the retransmission queue. The second number (3930) is the total number of retransmissions attempts.	count	0/3930	Indicates retransmit events & potential packet loss
Socket Buffer / Memory	rcv_space	Remaining receive buffer.The current amount of space available in the local receive buffer.	bytes	14,600	Prevents buffer overflow; flow control
	skmem	Total socket memory allocated	bytes	1,000,000	Overall socket memory usage
	rmem	Receive memory usage	bytes	60,000	Tracks receive buffer utilization
	wmem	Send memory usage	bytes	90,000	Tracks send buffer utilization
	oom	Out-of-memory events	count	0	Indicates memory exhaustion on socket
Misc / Kernel Internal	rxconf	RX queue / offload config	N/A	none	Debug RX offload / queue configuration
	txconf	TX queue / offload config	N/A	none	Debug TX offload / queue configuration
	tcp_ulp	Upper layer protocol (TLS offload)	N/A	tcp-ulp-tls	Shows TLS or other offload in use
	acked	Segments acknowledged	count	514	Confirms ACKed segments
	packets_out	Segments in flight	count	4	Indicates congestion / in-flight data
	sacked_out	SACKed segments pending ACK	count	2	Tracks SACKed segments for loss recovery
	fackets	Forward ACK segments	count	1	Tracks advanced ACK behavior for TCP reliability

 

Connection Info

ESTAB  0  71136  172.22.111.25:64413  172.25.21.1:34479

Field	Meaning
ESTAB	TCP connection state (ESTABLISHED).
0	Number of bytes queued in send buffer (unacknowledged but ready to send).
71136	Number of bytes in receive queue (data received but not read by application).
172.22.111.25:64413 → 172.25.21.1:34479	Source IP:port → Destination IP:port.

Congestion Control / TCP Parameters

cubic wscale:6,7 rto:670 backoff:1 rtt:134.721/0.928 ato:40 mss:1248 pmtu:1500 rcvmss:536 advmss:1460 cwnd:1 ssthresh:2

1. cubic
• Meaning: Congestion control algorithm used by this TCP connection.
• CUBIC: Default on modern Linux kernels. Optimized for high-speed, long-latency networks.
• Function: Controls how the congestion window (cwnd) grows/shrinks in response to network conditions.
• Effect: Determines throughput behavior during slow start, congestion avoidance, and recovery.
• Example: cubic → connection uses CUBIC algorithm rules for cwnd growth.

• Algorithm	Description	Typical Behavior
  CUBIC (default on modern Linux)	Non-linear congestion window growth optimized for high-BDP (Bandwidth-Delay Product) networks.	Grows slowly after congestion, then faster as time since last loss increases. Excellent for long fat networks (LFNs).
  Reno	Classic TCP algorithm; linear increase, multiplicative decrease.	Simple and fair but inefficient on high-latency or high-bandwidth paths.
  BBR (Bottleneck Bandwidth and RTT)	Estimates available bandwidth and minimum RTT to maximize throughput and minimize queueing delay.	Keeps queues short, high throughput, very responsive — often faster than CUBIC on clean links.
  BBRv2	Improved version of BBR with fairness improvements.	More fair against Reno/CUBIC, better coexistence.
  HighSpeed	Modified Reno for large bandwidth-delay networks.	Aggressive window growth, useful for data centers.
  Vegas	RTT-based algorithm (detects congestion before loss).	Keeps latency low, but less throughput on lossy links.
  Westwood+	Estimates bandwidth using ACK rate after packet loss.	Good for wireless/mobile networks.
  BIC	Predecessor of CUBIC; binary search window growth.	Replaced by CUBIC.

   
2. wscale:6,7
• Meaning: TCP Window Scale factor, applied to the advertised window to allow >64KB buffers.
• Format: wscale:<send scale>,<receive scale>
• 6 → sender window scale = 2⁶ = 64
• 7 → receiver window scale = 2⁷ = 128
• Effect: The actual window is TCP window field × 2^wscale.
• Why: Allows TCP to efficiently use large buffers on high-bandwidth, high-latency links.
• Example: rwnd of 501 in tcpdump × 128 → 64128 bytes effective receive window. 
3. rto:670
• Meaning: Retransmission Timeout (ms).
• Units: milliseconds
• Function: Time TCP waits before retransmitting an unacknowledged segment.
• Notes: RTO is adaptive, based on measured RTT + deviation (RTTvar).
• Effect: Too small → spurious retransmits; too high → slow recovery.
• Example: rto:670 → retransmit after 670 ms if no ACK received.
4. backoff:1
• Meaning: RTO backoff factor due to repeated timeouts.
• Function: TCP doubles RTO for each retransmission timeout (exponential backoff).
• Effect: Prevents flooding the network when persistent packet loss occurs.
• Example: backoff:1 → RTO currently multiplied by 2¹ = 2x (from base RTO).
• Explanation : TCP maintains a base RTO (Retransmission Timeout), computed dynamically from the RTT and its variation: RTO=SRTT+4×RTTVAR
  Let’s say your base (computed) RTO = 200 ms.
• First timeout happens, If an ACK isn’t received within 200 ms, TCP assumes packet loss and retransmits the unacknowledged segment.
  
  At this point: The connection experienced one timeout
  
  backoff = 1, TCP applies exponential backoff — meaning:
  
  Actual RTO=Base RTO×2*backoff
  
  So: RTO=200 ms×21=400 ms =400 ms
  
  That means after the first timeout, TCP will now wait 400 ms before retransmitting again if another ACK isn’t seen. 
• Another timeout (backoff increases again)
  
  If still no ACK after retransmitting, TCP increases backoff again → backoff = 2.
  
  Now: RTO=200 ms×22=800 ms
  
  And if yet another timeout occurs:
  
  backoff = 3 → RTO = 200 ms × 2³ = 1600 ms
• This is called exponential backoff — it protects the network during persistent loss or congestion by spacing out retransmissions, avoiding overload when the path is already unstable.

• Event	Base RTO (ms)	Backoff	Effective RTO (ms)	Behavior
  Initial transmission	200	0	200	Normal send
  1st timeout	200	1	400	2× delay
  2nd timeout	200	2	800	4× delay
  3rd timeout	200	3	1600	8× delay
  4th timeout	200	4	3200	16× delay

   
• Resetting backoff : Once a valid ACK is received and transmission stabilizes, backoff resets to 0 and RTO returns to its base (adaptive) value. backoff:1 means TCP has already experienced one retransmission timeout, so it has doubled the RTO to 2× its base value before trying again.
  
  
5. rtt:134.721/0.928
• Meaning: Measured Round Trip Time (RTT) and its variance.
• Units: milliseconds
• Format: <RTT> / <RTT variance>
• Effect: Used to compute adaptive RTO.
• Notes: High variance → more conservative RTO; low variance → faster retransmit.
• Example: rtt:134.721/0.928 → 134.721 ms average RTT, 0.928 ms variation.
• Explanation :  
• TCP measures RTT (Round Trip Time) — how long it takes for a packet to go to the receiver and get an ACK back.
• It also tracks RTT variance (how much the RTT fluctuates). 
• Then it calculates RTO (Retransmission Timeout) — the time TCP waits before retransmitting an unacknowledged packet. 
• The formula (RFC 6298) is roughly: RTO = SRTT + 4 × RTTVAR
• SRTT → Smoothed RTT (average RTT) 
• RTTVAR → Smoothed RTT variance 
• 🔹 Meaning of “High variance → more conservative RTO”
  
  If RTT values fluctuate a lot (network is unstable), then RTTVAR is high.
  → RTO becomes larger (since RTO = SRTT + 4×RTTVAR).
  → TCP waits longer before retransmitting.
  ✅ This prevents spurious retransmissions when ACKs are just delayed due to jitter.
• 🔹 Meaning of “Low variance → faster retransmit”
  
  If RTT values are stable and consistent, RTTVAR is small.
  → RTO becomes closer to SRTT (smaller timeout).
  → TCP retransmits faster when a packet is really lost.
  ✅ This improves throughput and responsiveness on stable links.

• Case	SRTT (ms)	RTTVAR (ms)	RTO (ms)	Behavior
  Stable network	100	2	108	Fast retransmit (tight timeout)
  Unstable network	100	30	220	Slow retransmit (safe buffer)

   
• Average RTT = 134.721 ms
  
  RTT variance = 0.928 ms (very low → stable network)
  → RTO will be ≈ 134.721 + 4×0.928 = ~138.4 ms
  → TCP will retransmit quickly if needed — since the connection is steady. 
6. ato:40
• Meaning: ACK timeout (delayed ACK timer) in ms.
• Function: Max time TCP waits before sending an ACK (to combine ACKs).
• Effect: Reduces small packet overhead but increases latency for small packets.
• Example: ato:40 → delayed ACK timer is 40 ms.
• Explanation : ato stands for ACK Timeout — the delayed ACK timer value in milliseconds. It represents how long TCP will wait before sending an ACK (acknowledgment) if no outgoing data is ready to "piggyback" the ACK on. 
• TCP can send an ACK immediately or delay it slightly to reduce overhead.
  The delayed ACK algorithm waits up to ato milliseconds, hoping that: More data arrives to acknowledge together, or The application sends some data back (so ACK can be piggybacked).This saves bandwidth and reduces small packets. 
• The TCP delayed ACK timer is 40 milliseconds.
• If the receiver gets a segment but has no data to send back, it will wait up to 40 ms before sending the ACK.
• If another packet arrives within that time, both can be acknowledged together. 
• Typical Linux Values :  sysctl -w net.ipv4.tcp_delack_min=20, sysctl -w net.ipv4.tcp_delack_max=40 

• Kernel	Default ato	Notes
  Older (2.6.x)	200 ms	High latency for small packets
  Modern (4.x–6.x)	40 ms	Tuned for better responsiveness
  Some distros	Adaptive (20–40 ms)	Based on RTT and application type

   
• Setting smaller values improves interactivity (e.g., in RPC or HTTPS handshakes), but increases ACK traffic. 
7. mss:1248
• Meaning: Maximum Segment Size negotiated for this connection.
• Units: bytes
• Function: Maximum TCP payload per segment, excluding headers.
• Effect: Defines largest chunk TCP will send per segment; smaller MSS → more segments, more overhead.
• Example: mss:1248 → each TCP segment carries max 1248 bytes of data.
8. pmtu:1500
• Meaning: Path MTU (Maximum Transmission Unit)
• Units: bytes
• Function: Maximum packet size that can traverse the path without fragmentation (IP + TCP headers included).
• Effect: Ensures packets aren’t dropped due to fragmentation.
• Example: pmtu:1500 → maximum full packet size = 1500 bytes.
9. rcvmss:536
• Meaning: Receiver’s maximum segment size (TCP advertised).
• Function: Sender must not send more than this per segment to avoid overwhelming receiver.
• Example: rcvmss:536 → sender must limit segments to 536 bytes payload.
10. advmss:1460
• Meaning: Peer’s advertised MSS.
• Function: Used by sender to segment outgoing data.
• Example: advmss:1460 → receiver can handle 1460 bytes per segment.

• Field	Who Sent It	What It Means	Impact on Server
  rcvmss:536	Client advertised this in its SYN	“I (Client) can receive at most 536 bytes per TCP segment.”	➜ Server must limit each outgoing segment to ≤ 536 bytes when sending data to this client.
  advmss:1460	Server advertised this in its own SYN-ACK	“I (Server) can receive up to 1460 bytes per TCP segment.”	➜ Client can send up to 1460-byte segments to the Server.

   
• It means:
  You (server) can receive 1460-byte segments. ✅
  You (server) must send smaller 536-byte segments to the client. ⚠️
• Possible Reason for rcvmss:536 The client is behind a low-MTU link (e.g., PPPoE, VPN, mobile, or older stack with 576-byte MTU). So, client advertised MSS=536 = 576 (MTU) - 40 (IPv4+TCP header). 
11. cwnd:1
• Meaning: Congestion window, sender-side.
• Units: MSS (segments)
• Function: Max number of segments allowed in flight (unacknowledged) at any time.
• Effect: Limits send rate; grows according to congestion control.
• Example: cwnd:1 → only 1 MSS in flight (connection is likely in slow start or recovering from loss).
12. ssthresh:2
• Meaning: Slow Start Threshold
• Units: MSS (segments)
• Function: Switch point from slow start (exponential growth) to congestion avoidance (linear growth).
• Effect: If cwnd < ssthresh → slow start; if cwnd ≥ ssthresh → congestion avoidance.
• Example: ssthresh:2 → after cwnd reaches 2 MSS, TCP will enter congestion avoidance. 

 

✅ Explanation of the timeline:

• Starts with cwnd=1 MSS → slow start doubles cwnd each RTT.
• Reaches ssthresh=2 MSS → growth becomes linear (congestion avoidance).
• Packet loss occurs → cwnd collapses to 1 (backoff), RTO waits 670ms.
• Retransmission resumes, cwnd grows again.
• MSS / PMTU limit the max bytes per segment; combined with cwnd, it limits in-flight data.
• RTT controls how fast ACKs arrive, pacing cwnd growth. 

Data Counters

bytes_sent:23793598 bytes_retrans:5059976 bytes_acked:18728630 bytes_received:331 segs_out:19069 segs_in:14656 data_segs_out:19067 data_segs_in:2

1. bytes_sent:23793598
• Meaning: Total number of payload bytes sent by this TCP connection since it was established.
• Units: Bytes (excluding TCP headers).
• Includes: Both successfully delivered bytes and retransmissions.
• Use: Shows total traffic volume originating from sender.
• Example: ~23.8 MB sent over this connection.
2. bytes_retrans:5059976
• Meaning: Number of bytes retransmitted due to packet loss or timeout.
• Units: Bytes
• Why it happens: 
• RTO (Retransmission Timeout) expired
• Triple duplicate ACK detected → fast retransmit
• Impact:
• Retransmissions indicate network congestion or packet loss
• Reduces effective throughput
• Directly affects cwnd (TCP reduces window on loss)
• Example: ~5.05 MB retransmitted → ~21% of total bytes sent. High value → poor network conditions.
3. bytes_acked:18728630
• Meaning: Number of bytes successfully acknowledged by the receiver.
• Units: Bytes
• Use:
• Shows effective throughput (how much data actually reached the peer)
• Helps compute loss rate: loss_rate ≈ bytes_retrans / bytes_sent ≈ 5.05M / 23.79M ≈ 21%
• Example: ~18.7 MB acknowledged → connection is sending more than what actually gets through without retransmission.
• Rule of thumb: <0.1–0.5% is fine; >1% is noticeable 
4. bytes_received:331
• Meaning: Total payload bytes received from peer (application layer data).
• Units: Bytes
• Notes:
• Very low here → sender is mostly transmitting; peer has sent minimal data.
• Could indicate a download-heavy or client-server upload-heavy scenario.
• Example: Only 331 bytes received → mainly upload connection.

• Metric	Formula	Healthy Range	Interpretation
  Loss rate	bytes_retrans / bytes_sent	< 0.5%	Higher = packet loss or congestion
  ACK efficiency	bytes_acked / bytes_sent	> 95%	Lower = retransmissions or stalled ACKs
  ACK coverage gap	bytes_sent - bytes_acked	small	Large gap = unacknowledged or inflight bytes
  TX/RX balance	bytes_sent / bytes_received	≈1 (full duplex) or ≫1 (upload)	Helps classify directionality

   

• Scenario	bytes_sent	bytes_retrans	bytes_acked	bytes_received	Interpretation
  💤 Idle	~0	~0	~0	~0	No active traffic; closed or paused connection.
  ✅ Healthy Active	High	Low (<0.5%)	Nearly = bytes_sent	Moderate/high	Normal flow, no loss or congestion.
  ⚠️ Mild Congestion	High	Moderate (1–2%)	Slightly below bytes_sent	Normal	Some retransmissions; cwnd adjusting.
  🚨 Severe Loss	High	High (>5%)	Much lower than bytes_sent	Normal	Network dropping packets; throughput collapsing.
  🧍 Unacknowledged / stalled	High	Growing	Stagnant	Normal	Peer not ACKing — maybe path blocked or flow control.
  🧾 Upload-heavy	High	Low–moderate	Matches sent	Very low	Sender uploads data; peer mostly ACKs (like file upload).
  📥 Download-heavy	Low	Low	Small	High	Receiving much more than sending; e.g., file download.
  🔁 Retransmission storm / RTO collapse	High	Extremely high (>20%)	Far below bytes_sent	Normal	Connection unusable; repeated losses.
  🧊 Receiver window full (rwnd=0)	Rising slowly	Low	Stalled	Stalled	Peer cannot read fast enough; flow control limiting speed.

   

• Derived Metric	Value	Interpretation
  Loss rate	5.05M / 23.79M = 21.2%	⚠️ Extremely high — massive retransmissions
  ACK efficiency	18.73M / 23.79M = 78.7%	Poor — 21% not ACKed successfully
  TX/RX ratio	≈ 71,800 : 1	Unidirectional — upload-only flow
  Summary	Server is mainly sending data; network has high packet loss or congestion.

   
5. segs_out:19069
• Meaning: Total number of TCP segments sent, including retransmissions.
• Includes: Data segments + pure ACKs + control segments.
• Units: Count of segments.
• Example: 19,069 segments sent over the connection.
6. segs_in:14656
• Meaning: Total number of segments received from peer.
• Includes: Data + ACKs + control packets.
• Units: Count of segments.
• Example: 14,656 segments received.
7. data_segs_out:19067
• Meaning: Number of segments carrying actual TCP payload sent.
• Difference vs segs_out:
• segs_out = all segments including pure ACKs
• data_segs_out = only segments with real data
• Example: 19,067 segments carry actual payload → 2 segments may have been pure ACKs or control.
8. data_segs_in:2
• Meaning: Number of segments carrying actual payload received from peer.
• Notes: Very low → peer is barely sending data.
• Example: 2 segments → peer sends al

 

Throughput / Timing

send 74.1kbps pacing_rate 356kbps delivery_rate 220kbps lastsnd:655 lastrcv:1591060 lastack:1266 busy:1592118ms

1. send 74.1kbps
• Meaning: Current instantaneous send rate of this TCP connection.
• Units: kilobits per second (kbps)
• How it’s computed:
• Based on the number of bytes actually transmitted over the recent RTT
• Only includes new transmissions, not retransmissions.
• Observation in your case:
• 74.1 kbps is very low → connection is congestion-limited, probably due to cwnd=1 and retransmissions.
2. pacing_rate 356kbps
• Meaning: TCP pacing rate, used if Linux TCP pacing is enabled.
• Units: kbps
• Function: Limits how fast the sender injects packets into the network to avoid bursts.
• Observation:
• Kernel allows up to 356 kbps, but actual send = 74.1 kbps → sending slower due to cwnd/retransmissions.
3. delivery_rate 220kbps
• Meaning: Measured rate of successful delivery of bytes to the receiver.
• Units: kbps
• Difference from send/pacing:
• send = current bytes leaving kernel
• delivery_rate = how fast data is ACKed and confirmed delivered
• Observation:
• 220 kbps → actual effective throughput higher than instantaneous send?
• This may include bursts or averaged over longer RTT window.
•  

  Scenario	send	pacing_rate	delivery_rate	Diagnosis
  ✅ Healthy steady state	≈ pacing_rate	≈ pacing_rate	≈ pacing_rate	Fully utilizing allowed bandwidth
  ⚠️ Loss recovery / cwnd shrink	≪ pacing_rate	steady or reduced	< pacing_rate	Send rate throttled by cwnd, not pacing
  🧊 ACK delay or high RTT	≈ pacing_rate	high	lower	Data sent quickly, ACKs arrive slowly
  🧱 Severe congestion / RTO	tiny	larger	tiny	Connection nearly stalled
  💤 Idle / app limited	0	steady	0	Application not producing data

•  Quick Diagnostic Guide

• Pattern	Likely Cause	Typical Action
  send ≈ pacing_rate ≈ delivery_rate	Normal / balanced	None
  send << pacing_rate	cwnd too small (loss recovery)	Investigate bytes_retrans, RTT spikes
  delivery_rate << send	ACK loss or receiver slow	Check rtt, rttvar, rwnd
  All near-zero	Idle or app paused	App-limited, normal

   
•  In Above Numbers (74 kbps vs 356 kbps vs 220 kbps)

• Observation	Implication
  send << pacing_rate	Sender is holding back — likely due to small cwnd after losses (remember you had ~21 % retransmits).
  delivery_rate > send	Receiver ACKs older bursts; temporary effect of delayed ACK catching up.
  Combined →	Connection is congestion-limited, not pacing-limited. Linux pacing permits 356 kbps, but cwnd or RTO recovery restricts actual send to ~74 kbps.

   
4. lastsnd:655
• Meaning: Timestamp of last data sent on this connection (ms since connection start).
• Effect: Useful to detect idle periods or pacing behavior.
5. lastrcv:1591060
• Meaning: Timestamp of last segment received from peer (ms since connection start).
• Observation: Huge difference vs lastsnd → peer is sending almost nothing (matches bytes_received=331).
6. lastack:1266
• Meaning: Timestamp of last ACK received from the peer.
• Function: Determines whether unacknowledged data exists and affects cwnd growth.
7. busy:1592118ms
• Meaning: Total time this TCP connection has been active in kernel (microseconds → ms).
• Observation: Connection has been active ~1,592 seconds (~26 minutes). 

 

Insights

• Send vs Pacing vs Delivery:
• The kernel can send faster (pacing_rate), but cwnd = 1 → throttles actual send rate.
• Delivery rate is higher than instantaneous send due to averaging or bursts.
• Low send rate:
• Matches earlier observation of cwnd=1 and high retransmissions (bytes_retrans).
• Idle detection:
• lastsnd vs lastrcv → peer sending almost nothing → connection mostly upload.
• Performance bottleneck:
• Not NIC or link capacity (pacing_rate=356 kbps), TCP congestion control and retransmissions are limiting throughput. 

 Queue / Retransmission Info

unacked:4 retrans:1/4055 lost:4 reordering:20 rcv_space:14600 rcv_ssthresh:64076 notsent:66144 minrtt:127.84 snd_wnd:76096

1. unacked:4
• Meaning: Number of segments sent but not yet acknowledged by the receiver.
• Units: Segments (MSS units)
• Impact:
• Determines how much data is currently “in flight.”
• If unacked >= cwnd, TCP cannot send more until ACKs arrive.
• Example: 4 segments unacknowledged → limits sending if cwnd is small (matches cwnd=1 from earlier).
2. retrans:1/4055
• Meaning: Retransmissions in two forms:
• 1 → currently scheduled for retransmission
• 4055 → total segments retransmitted so far
• Cause: Packet loss detected via RTO or triple duplicate ACKs
• Impact:
• Each retransmission reduces effective throughput
• Triggers congestion control (reduces cwnd)
• Example: 1 segment pending retransmit; 4055 already retransmitted → high-loss path.
3. lost:4
• Meaning: Segments marked lost by TCP (timeout or duplicate ACK detection).
• Impact:
• TCP reduces cwnd (usually to 1 or half of ssthresh)
• Triggers retransmissions
• Example: 4 segments lost → explains why cwnd=1 and send rate is low.
4. reordering:20
• Meaning: Number of segments received out-of-order.
• Impact:
• Can trigger duplicate ACKs → potential fast retransmit
• High reordering can simulate packet loss in TCP perception
• Observation: 20 out-of-order segments → moderate reordering, typical on multi-path or virtual networks.
5. rcv_space:14600
• Meaning: Remaining receiver buffer space in bytes.
• Impact:
• Limits how much data the sender can push
• Effective send window = min(cwnd × MSS, snd_wnd)
• Observation: Receiver has ~14.6 KB free buffer → not currently the bottleneck.
6. rcv_ssthresh:64076
• Meaning: Receiver-side slow start threshold (in bytes)
• Impact: Guides flow control on how quickly the receiver can accept data
• Observation: 64 KB threshold → receiver can handle bursts before signaling congestion.
7. notsent:66144
• Meaning: Bytes in the send buffer waiting to be transmitted (not yet sent).
• Impact:
• Limited by cwnd and pacing
• Indicates backpressure in sender queue
• Example: 66 KB waiting → can’t send immediately because cwnd is too small.
8. minrtt:127.84
• Meaning: Minimum RTT observed on this connection (ms)
• Impact:
• Used by congestion control (CUBIC/TCP) to estimate bandwidth
• Helps compute pacing and retransmission timing
• Example: 127 ms → baseline RTT for cwnd growth calculations.
9. snd_wnd:76096
• Meaning: Sender’s view of receiver advertised window (rwnd)
• Units: Bytes
• Impact:
• TCP can send at most min(cwnd × MSS, snd_wnd)
• rwnd limits maximum in-flight data
• Example: 76 KB → receiver window is sufficient; not limiting in-flight data here. 

 

Key Insights

• Sender congestion-limited: cwnd=1 + 4 unacked → can't send all 66 KB in buffer.
• High retransmissions & lost segments → explains low send rate (74 kbps).
• Receiver window sufficient: snd_wnd=76 KB > cwnd*MSS → not limiting.
• Reordering moderate: could trigger extra duplicate ACKs, affecting cwnd.
• Min RTT = 127 ms: sets baseline for CUBIC growth; pacing may wait for ACKs. 

TLS / ULP Info

tcp-ulp-tls rxconf: none txconf: none

Field	Meaning
tcp-ulp-tls	Using TLS as a TCP Upper Layer Protocol (kernel-level TLS offload).
rxconf / txconf	TLS configuration on RX/TX side (none configured here).