Benchmarking CurveCP Thorben Kr uger benthor@os3.nl July 4, 2011 - PowerPoint PPT Presentation

Benchmarking CurveCP Thorben Kr¨ uger benthor@os3.nl July 4, 2011 1/ 30

Outline Introduction Research Questions Methods and Results Problems Conclusion 2/ 30

CurveCP? Encrypted application-layer protocol for internet communication. 3/ 30

CurveCP in a Nutshell ◮ Brain child of Dan Bernstein (djb) ◮ To be used instead of TCP ◮ Packet-based encryption on top of UDP ◮ Treat crypto as instantaneous ◮ Advanced packet scheduler 4/ 30

Claimed CurveCP Security Features ◮ mandatory server authentication ◮ optional client authentication ◮ no man-in-the-middle attacks possible ◮ active and passive forward secrecy 5/ 30

Claimed CurveCP Availability Features ◮ no RST-type attacks possible ◮ protection against traffic prediction ◮ can not be used for amplification attacks ◮ no SYN-flooding-type attacks possible ◮ worst-case CPU loads kept small 6/ 30

Claims about CurveCP Efficiency ◮ bigger overhead than plain TCP ◮ for short connections, less traffic than HTTPS ◮ for short connections, much less traffic than SSH 7/ 30

Claimed CurveCP Decongestion Features ◮ minimizes packet-loss ◮ minimizes significant latency increases ◮ therefore mitigates buffer bloat 8/ 30

Claimed CurveCP Addressing Features ◮ multiple CurveCP servers can share single IPv4 address and port ◮ CurveCP servers inherently anti-aliased from addresses ◮ rapid failover to redundant server if original is down ◮ session/connection not invalidated if IP address changes 9/ 30

Major Topics for Investiagtion Create CurveCP-enabled SSH-like remote shell Create CurveCP-enablet SCP-like remote file copy tool Benchmark CurveCP vs SSH/SCP/HTTPS ◮ CPU usage ◮ available Bandwidth ◮ ideal ◮ competetive ◮ message latencies Verify CurveCP robustness claims 10/ 30

Remote Shell/Copy via CurveCP Results: ◮ CurveCP-enabled remote ptty fully functional ◮ http://github.com/benthor/remotty Problems: ◮ Only Python implementation so far ◮ Useless for meaningful benchmarks ◮ No real session handling yet ◮ Doesn’t support file transfer yet 11/ 30

CurveCP-enabled file transfer? For now based on cat / dd ◮ No support for arbitrary files yet 12/ 30

Bonus: CurveCP enabled VPN Results: ◮ Can create tunnel devices connected via CurveCP ◮ ICMP can be successfully tunneled Problems: ◮ TCP only works for small packets so far ◮ Only Python implementation 13/ 30

Benchmarking: Methods Tools: ◮ plain dd ◮ custom stream copy tool with statistics ( ddstat ) ◮ sysstat suite ◮ ethtool 14/ 30

Saturating 10Mbit Link Results: ◮ CPU usage of SSH/SCP/HTTPS very similar Problems: ◮ CurveCP scheduler is too nice 15/ 30

Best-case Payload Bandwidth on 10MBit Link Results: ◮ SSH/SCP/HTTPS: nearly the full 10MBit/s ◮ CurveCP: between 600KB/s and 800KB/s, average 650KB/s Problems: ◮ CurveCP scheduler is too nice 16/ 30

CurveCP and TCP: Competitive Scheduling (10MBit Link) Problems: ◮ CurveCP bandwidth drops to 0 ◮ no matter who starts first ◮ no matter if LAN or Internet connection 17/ 30

Trying to saturate 100Mbit Link Results: ◮ SSH/SCP/HTTPS can saturate ◮ CurveCP probably CPU bound Problems: ◮ Detailed measurements skew CPU usage 18/ 30

Best-case Payload Bandwidth on 100MBit Link Results: ◮ SSH/SCP/HTTPS: nearly the full 12.5MB/s ◮ CurveCP: tops out at 4MB/s 19/ 30

CurveCP and TCP: Competitive Scheduling (100MBit Link) Results: ◮ CurveCP bandwidth 1/100th of that of TCP: 120KB/s ◮ More than one connection: share up to 240KB/s 20/ 30

CurveCP vs SSH: Latencies Results: ◮ (Extremeny) similar latencies to within fractions of ms Problems: ◮ Does not take SSH handshake into account 21/ 30

Bonus: CurveCP and Buffer Bloat Queue delay on top of minimum path RTT - bufferbloat-curvecp1 1400 1200 Added Queue delay(ms) 1000 800 600 400 200 0 J J J J J J J J J J J J J J J J J J J u u u u u u u u u u u u u u u u u u u n n n n n n n n n n n n n n n n n n n 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 8 8 8 8 9 9 9 9 9 9 9 9 9 9 9 9 0 0 0 : : : : : : : : : : : : : : : : : : : 4 4 5 5 0 0 1 1 2 2 3 3 4 4 5 5 0 4 4 3 8 3 8 3 8 3 8 3 8 3 8 3 8 3 8 3 2 7 Time Min path Link 1 Link 2 Link 3 Link 4 Link 5 Link 6 22/ 30

Questions not yet answered Will be addressed in paper: ◮ Compare SSH handshake to CurveCP ◮ CurveCP overhead in general ◮ CurveCP addressing and failure modes 23/ 30

Encountered Problems 24/ 30

Problem: (Reverse) HeisenBugs Packet scheduler gets confused: ◮ when UDP statistics are collected ◮ when in the presence of irregular TCP traffic Result: ◮ endless tracing/profiling/code-reviewing 25/ 30

Problem: FreeBSD ◮ CurveCP pipes close prematurely ◮ file descriptor issues 26/ 30

Problem: Confusing CPU statistics ◮ disagreement between tools ◮ reported percentages add up to over 100 27/ 30

Conclusion ◮ Remote shell: works ◮ Remote copy: possible ◮ Performance: comparable to (but worse than) SSH/HTTPS ◮ Decongestion: works somewhat ◮ Verdict: CurveCP worthy of attention 28/ 30

More Info http://curvecp.org 29/ 30

Questions? 30/ 30