| Network Working Group | J. Mattsson |
| Internet-Draft | Ericsson |
| Intended status: Informational | October 27, 2014 |
| Expires: April 30, 2015 |
Overview and Analysis of Overhead Caused by TLS
draft-mattsson-uta-tls-overhead-01
A common argument against the use of TLS is that it adds overhead. In this document we illustrate in detail how much (or little) processing, latency, and traffic overhead TLS adds. Transition to more secure cipher suites (TLS 1.2 with AES-GCM or ChaCha20-Poly1305) actually reduces both traffic and processing overhead. AES-GCM combines security, low traffic overhead, and great performance on modern hardware. On platforms without hardware support for AES-GCM, ChaCha20-Poly1305 gives the same benefits. For everything but very short connections, TLS is not inducing any major traffic overhead (nor CPU or memory overhead).
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at http://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on April 30, 2015.
Copyright (c) 2014 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
The overhead from TLS can be divided into several different aspects:
But in many scenarios, TLS does not add much overhead at all, and moving to more secure cipher suites actually reduces both traffic and processing overhead.
The TLS handshake typically adds 4-7 kB of traffic overhead. TLS compression reduces traffic overhead, but has negative security implications and should be turned off [I-D.ietf-uta-tls-bcp].
Looking at the certificates, a move from 1024 to 2048 bit RSA keys increases traffic and processing overhead but is needed for security reasons. Certificates with 1024 bit RSA keys should be phased out as they only gives 80 bit security. NIST recommendation is to stop using algorithms giving 80 bit security no later that 2010 [KeyLength]. A move from SHA-1 to SHA-256 adds processing overhead but is needed for security reasons. SHA-1 should not be used anymore for digital signatures (e.g. in certificates) as it gives less than 80 bit security. To summarize, SHA-2 certificates with at least 2048 bit RSA keys should be used.
In TLS 1.2 [RFC5246] and earlier versions, the initial handshake takes 2 round-trips and session resumption takes 1 round-trip. In TLS 1.3 [I-D.ietf-tls-tls13] the target is 1 round-trip for the initial round-trip and 0 round-trips for session resumption. Because of the emphasis on reducing latency (instead of only security), TLS 1.3 is expected to have much faster deployment than earlier versions.
Some of the most commonly used ciphersuites have security weaknesses. Encryption algorithms such as RC4 and the CBC modes (e.g. AES and 3DES_EDE) have security weaknesses, and the hash functions SHA-1 and MD5 (but not the HMAC constructions used in TLS record layer) also have security weaknesses.
More recent ciphersuites using AES-GCM and CHACHA20_POLY1305 have no known security weaknesses, but AES-GCM, CHACHA20_POLY1305 and other AEAD suites requires TLS 1.2 [RFC5246]. CHACHA20_POLY1305 is currently only an Internet draft but is still used in practice as it is very fast in software [I-D.agl-tls-chacha20poly1305]. AES-GCM is the current IETF recommendation (Internet Draft) as part of TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256 [I-D.ietf-uta-tls-bcp].
These data is included as to motivate which algorithms to cover as well as showing that new secure ciphersuites are gaining signifiacant usage. In data from July 2014 [ICSI], AES-CBC, RC4, and HMAC-SHA1 dominates, these ciphersuites all have security weaknesses. The NULL cipher does not provide any confidentiality at all. The more secure options AES-GCM and ChaCha20-Poly1305 are starting to showing significant usage. 3DES_EDE_CBC_SHA is includes as it is mandatory to implement in TLS 1.0.
Algorithm Usage ---------------------------------------------------------- AES_128_CBC_SHA 29.1 % RC4_128_SHA 17.4 % AES_128_GCM 14.7 % AES_256_CBC_SHA 14.0 % NULL_SHA 9.8 % RC4_128_MD5 8.3 % CHACHA20_POLY1305 1.4 % 3DES_EDE_CBC_SHA < 5.6 % ----------------------------------------------------------
Table 1: Ciphers in Use (ICSI, July 2014)
The traffic overhead comes in different forms: the TLS record layer header, Explicit IV/Nonce, MAC tag, and encryption algorithm padding. Figure 1 illustrates the packet format for a TLS protected package where [ ] indicates fields where usage depends on the TLS version and the ciphersuite used.
+----+-----+------------+------------+--------------+-----+-----------+ | IP | TCP | TLS Header | [IV/Nonce] | Enc. Content | MAC | [Padding] | +----+-----+------------+------------+--------------+-----+-----------+
Figure 1: Format of TLS protected packet
The size of the TLS header is fixed (5 bytes). The size of the IV/Nonce depends on the TLS version and the ciphersuite used. Explicit IV is used by CBC ciphersuites in TLS 1.1 and TLS 1.2, but not TLS 1.0. Explicit Nonce is used by AEAD algorithms. The size of the MAC tag depends on the ciphersuite used, it is either a separate field (non-AEAD algorithms like SHA-1, MD5) or included in the ciphertext (AEAD algorithms like GCM, Poly1305). Padding is used by CBC ciphersuites.
The per-packet overhead for the most important ciphersuites are shown below (the values are all theoretical and the averages are calculates over a uniform distribution). For comparision, the TCP/IP overhead for IPv4 and IPv6 are 52 and 72 bytes, respectively.
----------------------------------------------------------------- AES_128_CBC_SHA, AES_256_CBC_SHA ----------------------------------------------------------------- Per-packet overhead (TLS 1.0) 26-41 bytes (avg. 33.5) TLS header 5 bytes HMAC-SHA-1 20 bytes CBC padding 1-16 bytes Per-packet overhead (TLS 1.1, 1.2) 42-57 bytes (avg. 49.5) TLS header 5 bytes Explicit IV 16 bytes HMAC-SHA-1 20 bytes CBC padding 1-16 bytes -----------------------------------------------------------------
----------------------------------------------------------------- 3DES_EDE_CBC_SHA ----------------------------------------------------------------- Per-packet overhead (TLS 1.0) 26-33 bytes (avg. 29.5) TLS header 5 bytes HMAC-SHA-1 20 bytes CBC padding 1-8 bytes Per-packet overhead (TLS 1.1, 1.2) 34-41 bytes (avg. 37.5) TLS header 5 bytes Explicit IV 8 bytes HMAC-SHA-1 20 bytes CBC padding 1-8 bytes -----------------------------------------------------------------
----------------------------------------------------------------- RC4_128_SHA, NULL_SHA ----------------------------------------------------------------- Per-packet overhead (TLS 1.0, 1.1, 1.2) 25 bytes TLS header 5 bytes HMAC-SHA-1 20 bytes -----------------------------------------------------------------
----------------------------------------------------------------- RC4_128_MD5 ----------------------------------------------------------------- Per-packet overhead (TLS 1.0, 1.1, 1.2) 21 bytes TLS header 5 bytes HMAC-MD5 16 bytes -----------------------------------------------------------------
----------------------------------------------------------------- AES_128_GCM, AES_256_GCM ----------------------------------------------------------------- Per-packet overhead (TLS 1.0, 1.1, 1.2) 29 bytes TLS header 5 bytes Explicit Nonce 8 bytes GMAC 16 bytes -----------------------------------------------------------------
----------------------------------------------------------------- CHACHA20_POLY1305 ----------------------------------------------------------------- Per-packet overhead (TLS 1.0, 1.1, 1.2) 29 bytes TLS header 5 bytes Explicit Nonce 8 bytes Poly1305 16 bytes -----------------------------------------------------------------
As can be seen from the tables above, there is a correlation between better security and low traffic overhead. Going from TLS 1.1 [RFC4346] with AES_CBC_SHA (mandatory to implement in TLS 1.1) to TLS 1.2 [RFC5246] with one of the more secure options AES_GCM (current IETF recommendation) or CHACHA20_POLY1305 reduces record layer traffic overhead with 41 %. Going from TLS 1.0 [RFC2246] with AES_CBC_SHA to TLS 1.2 with AES_GCM or CHACHA20_POLY1305 reduces record layer traffic overhead with 14 %.
Just as with traffic overhead, there is a correlation between better security and low processing overhead. Going from AES_CBC_SHA (mandatory to implement in TLS 1.1. and 1.2) to the more secure option AES-GCM reduces processing overhead on a Core-i7-3770 processor with 57 %. Another fact is that the overhead for AES_128_GCM and CHACHA20_POLY1305 is so low, there is no overhead reasons to not use encryption (i.e. NULL_SHA).
On modern x86 processors with hardware support for AES (AES-NI) and carry-less multplication (CLMUL), AES_GCM is much faster that RC4_SHA, AES_CBC_SHA, or CHACHA20_POLY1305. Another performance advantage with AES-GCM is that is designed for parallelization.
Algorithm Speed ---------------------------------------------------------- AES_128_GCM 1909.1 MB/s CHACHA20_POLY1305 625.2 MB/s AES_128_CBC_SHA 573.7 MB/s AES_256_CBC_SHA 486.6 MB/s RC4_128_MD5 233.9 MB/s ----------------------------------------------------------
Table 2: Speed on 2 GHz Intel Core i7
These measuments are not fair to Chacha20-Poly1305, but this does not matter, the important thing is who fast does the algorithms run on current hardwware.
Without hardware support for AES-GCM, ChaCha20-Poly1305 is much faster that AES-GCM (and AES-CBC). Data from [Software].
Algorithm Speed ---------------------------------------------------------- CHACHA20_POLY1305 130.9 MB/s AES_128_GCM 41.5 MB/s ----------------------------------------------------------
Table 3: Speed on Snapdragon S4 Pro
Several companies have deployed ChaCha20-Poly1305 to get better performance (and security) on platforms without AES and CLMUL hardware support. This may have less significance in the future if mobile CPUs implement hardware support for AES-GCM.
Transition to more secure cipher suites (TLS 1.2 with AES-GCM or ChaCha20-Poly1305) actually reduces both traffic and processing overhead. Going from TLS 1.1 with AES_CBC_SHA (mandatory to implement in TLS 1.1) to TLS 1.2 with AES_GCM (current IETF recommendation) or CHACHA20_POLY1305 reduces record layer traffic overhead with 41 %, and record layer processing overhead with even more. AES-GCM combines security, low traffic overhead, and great performance on modern x86 hardware. On platforms without hardware support for AES-GCM, ChaCha20-Poly1305 gives the same benefits.
Looking at the certificates, a transition to SHA-2 certificates with RSA-2048 keys increases TLS handshake traffic and processing overhead but is needed for security reasons.
For everything but very short connections, TLS is not inducing any major traffic overhead (nor CPU or memory overhead). Server people from Google Gmail has stated that “TLS accounts for less than 1% of the CPU load, less than 10 KB of memory per connection and less than 2% of network overhead”. Main impact of TLS is increased latency, this can by reduced by using session resumption, cache information closer to end users, or waiting for TLS 1.3.
The whole document is about increasing the use of TLS and secure ciphersuites by showing that TLS is many cases does not add much overhead, and that there for many types of overhead is a correlation between better security and low overhead.
The authors would like to thank Stephen Farrell and Ivan Ristic for their valuable comments and feedback.