All research
10 min readLatencyMicrostructure

Cross-venue L1 latency from Railway SG, US-East, US-West

What a five-minute, three-region measurement across eleven venues says about cross-reference feed selection for a perp-DEX market-making operation.

A useful cross-venue reference signal has to arrive at your decision loop faster than the target venue's own feed — otherwise you're predicting the past. We ran a three-region measurement exercise to understand which venues can serve as cross-reference feeds for a crypto-derivatives market-making operation deployed on Railway, and how that ranking shifts across regions.

TL;DR

  • Singapore is the best region we tested. Bybit's top-of-book arrives in 2–3 ms p50 — an order of magnitude faster than the next venue from Singapore, and ~49 ms ahead of a typical perp-DEX own-feed latency.
  • US-West is a viable Plan B. Gate and Binance land at ~54 ms p50; the lead over a perp-DEX own-feed is 10–15 ms. Smaller, still real.
  • US-East has the largest apparent lead on paper but also the slowest order path to the target venue (~147 ms). By the time a reaction order lands, the lead has eroded. Not recommended.
  • Aster is the tightest-distribution DEX in every region (p95 within 5–10 ms of p50). Unusual among DEXs — most behave more like Hyperliquid.
  • Lighter is observable only from Singapore on datacenter infrastructure; Extended has a bimodal tail (p95 13–40 s) everywhere and is unusable as a real-time signal.

Methodology

One child process per venue, subscribed to top-of-book for six tokens (BTC, ETH, SOL, XRP, BNB, HYPE). Each sample records (venue, token, exchange_timestamp_ms, local_recv_wall_ms) and latency is computed against the venue-published event-time field (Binance E, OKX/Bybit ts, Gate result.t, and so on).

Eleven venues were sampled: Binance, Bybit, OKX, KuCoin, Gate, Aster, Hyperliquid, Lighter, Extended, EdgeX, and a target perp DEX (GRVT). Each region was run for five minutes with the same universe, captured on Railway at asia-southeast1, us-east4, and us-west2. NTP sync was clean across all three runs; zero negative-latency samples.

Five minutes is short for tail statistics but enough for medians and p95 within ~20 ms. Absolute maxima are indicative, not authoritative.

Singapore

From asia-southeast1, Bybit's latency is extraordinary: 2–3 ms p50 across every token, p95 within 1–2 ms of p50. No other venue in the universe comes within 10× on median. KuCoin (38 ms), Binance (39–48 ms), Gate (39 ms), and Aster (42 ms) form the second tier — all Tokyo-/Asia-routed, all with tight distributions. OKX at 63 ms p50 was ~3× slower than a prior measurement from the same region; we flagged the instability but didn't track it down.

Venuep50 (ms)p95 (ms)Notes
Bybit2–33–4Uniquely fast from SG
KuCoin3838–45Trade-driven, sparse
Binance39–4892–117Tokyo-routed
Gate3940–59Tight, consistent
Aster4245–48Tightest DEX distribution
GRVT (target)49–71307–322Own-feed reference
OKX6363–65Instability vs prior runs
Hyperliquid239–254545–610Block-time-bound

US-East

From us-east4, KuCoin (71 ms), Gate (80 ms), Binance (82–84 ms), and Aster (83 ms) form the fastest tier. Bybit (107 ms) loses its Singapore advantage decisively. GRVT's own feed is unusually slow from US-East — ~168 ms p50, roughly 3× its latency from the other two regions — suggesting GRVT's infrastructure footprint doesn't reach the US east coast.

The large apparent lead (Gate beats GRVT's own feed by 88 ms) is misleading in practice: the order path from US-East to GRVT is itself ~147 ms, so by the time a reaction order lands the apparent edge has been consumed.

US-West

Railway us-west2 is the middle ground. Gate (54 ms), Binance (54–56 ms), KuCoin (56 ms), and Aster (57–58 ms) tie at roughly 54–58 ms p50. GRVT's own feed is 64–77 ms p50 — the same relative ordering as Singapore, but compressed. Lead over the own-feed is ~10–15 ms: smaller, still positive. Hyperliquid is fastest from US-West (206 ms p50), suggesting an HL backend with US-West bias — still too slow to serve as a cross-reference for a target venue whose own feed arrives in ~70 ms.

Cross-region synthesis

VenueSG p50US-East p50US-West p50Interpretation
Bybit310784Singapore-local
Gate398054Fastest CEX from US-West
Binance39–4882–8454–56Tokyo; SG ≈ US-West
Aster428357–58Tightest-distribution DEX
GRVT (target)49–71164–17264–77Infra US-West-leaning
OKX6310881
Hyperliquid239–254287–314206–217Block-time-bound

The three-region picture:

  • Bybit is region-conditional. From Singapore it is uniquely fast; from US-West it is comparable to everyone else; from US-East it is slower than several cheaper candidates. “Use Bybit as your cross-reference” is a region-conditional claim, not a venue-intrinsic one.
  • Gate is the most consistently fast CEX across regions. Never the absolute winner, always in the top tier.
  • Aster has CEX-like distribution stability. p95 sits within 5–10 ms of p50 for every token in every region. Among perp DEXs that is unusual — most behave more like Hyperliquid.
  • GRVT's infrastructure is US-West-leaning. Own-feed latencies cluster as SG ≈ US-West < US-East.
  • Hyperliquid is block-time-bound, not location-bound. The ~200 ms floor reflects HL's consensus interval, not network distance. Co-location doesn't meaningfully help.

Known broken venues

  • Lighter fails from any US region. WebSocket handshake returns HTTP 400; 14 reconnects inside five minutes from both US-East and US-West; clean from Singapore. Root cause is almost certainly a CloudFront WAF rule that blocks cloud-provider ASNs at the US PoPs. Not fixable client-side without a proxy or non-datacenter IP. Lighter is effectively only observable from Singapore on infrastructure like Railway.
  • Extended has a bimodal tail everywhere. p50 is reasonable (50–130 ms), but p95 is 13–40 seconds and max is 38–59 seconds. The pattern looks like periodic server-side buffering and burst delivery. Unusable as a real-time reference until it's fixed upstream.
  • EdgeX produced one sample per token across 18 independent trials (3 regions × 6 tokens × 5 minutes). Client-side bug; under investigation.

Operational conclusion

For a market-making operation on a perp DEX reachable from Railway, where the reference target's own feed runs ~50 ms p50:

RankRegionCross-ref venueLead over own-feedOrder pathViable?
1SGBybit~49 ms~43 msYes — recommended
2US-WestGate / Binance~10–15 ms~56 msMarginal — Plan B
3US-EastGate~88 ms (apparent)~147 msNo — order path kills it

What this does not tell you

Latency is an input; it is not an edge by itself. The realised value of a cross-venue signal depends on lead-lag correlation, the mapping between venues' book representations, and cost structure at the target venue. Even a 49 ms absolute edge collapses to zero if the reference and target venues price off different underlyings, or if the target venue's microstructure doesn't respond to cross-venue moves in the way the signal predicts.

We would also caution against reading single-run tail figures too strictly. Five minutes is enough to characterize medians; daily-cadence monitoring is the next step if stability matters.