Responsiveness diagnostic

Bufferbloat test

Find out whether your connection stays responsive while it is busy. Compare idle latency with separate download-loaded and upload-loaded medians, p95 tail delay, jitter, and browser HTTP probe misses.

Idle + loaded latencyDownload and upload separatedRegional endpointsNo account

Data-use notice: the tester starts in full-capacity mode so it can create meaningful load. Review the displayed transfer estimate before starting and choose low-data mode if the connection is metered.

Test latency under load

Compare idle latency with download-loaded and upload-loaded latency using the full-capacity test.

Goal

Connection

Provider

Place

Optional context is saved with recent results and included only if you copy a share link.

Full-capacity measurement

Active because you selected full-capacity mode. The time- and byte-bounded run requests at most about 832 MiB of test payload, and usually less. Protocol overhead may add a little more.

Test endpoint

Default Server - Nearest Location

Same-origin hosted endpoint

Endpoint readiness unavailable

Default endpoint is ready.

For a controlled baseline, pause unrelated traffic; this test creates its own download and upload load.

Transparent grading

Exact latency-under-load grade boundaries

This is a SwiftSpeedTest editorial diagnostic, not a universal industry certification. The run receives the first grade for which every boundary is met. The worst download or upload direction is used; anything beyond every D boundary receives F.

Grade	Median rise	p95 tail rise	Loaded jitter	Discarded samples	HTTP probe misses
A	≤ 15 ms	≤ 40 ms	≤ 10 ms	< 5%	< 1%
B	≤ 30 ms	≤ 75 ms	≤ 20 ms	< 10%	< 2%
C	≤ 60 ms	≤ 150 ms	≤ 40 ms	< 20%	< 5%
D	≤ 120 ms	≤ 300 ms	≤ 80 ms	< 35%	< 10%
F	One or more metrics exceed every D boundary.

Discarded samples are successful but statistically extreme values excluded from the retained median; they remain visible through the discarded ratio and p95 diagnostics. HTTP probe misses are failed or timed-out application requests, not network-layer packet loss.

What the result means

Capacity and responsiveness are different

A connection can deliver high Mbps while interactive traffic stalls behind a long queue. The useful comparison is the change from idle latency to latency while download and upload traffic are active—not the loaded value by itself.

Download rise: queueing while the connection receives sustained traffic.
Upload rise: queueing while backups, calls, or other traffic sends data.
p95 tail: the slower end of the successful probe series, preserved so a stable median does not hide repeated stalls.
HTTP misses: unsuccessful application probes that must not be relabeled as ICMP or UDP packet loss.

Controlled diagnosis

Isolate the queue before changing settings

1. Repeat once. Keep the endpoint, device, and connection type fixed to rule out a single noisy run.
2. Compare Ethernet and Wi-Fi. A large wireless-only change points toward the local radio, driver, or access point rather than proving an ISP issue.
3. Check the worse direction. Upload-only inflation often appears during cloud backup, camera sync, or large sends; download-only inflation needs a different controlled reproduction.
4. Evaluate documented queue management. If the router supports SQM, FQ-CoDel, CAKE, CoDel, or PIE, follow that device's instructions and retest after one change. Do not assume a label alone proves the right setting.

For intermittent problems without intentional load, use the internet stability test.

Methodology sources

These references define bufferbloat, latency under load, and queue-management concepts. SwiftSpeedTest applies them to a versioned browser HTTP diagnostic and separately publishes its own grading boundaries. References checked 2026-07-24.

IETF: RFC 7928: Characterization Guidelines for AQM
Describes bufferbloat as excessive unmanaged buffering that increases end-to-end delay and harms latency-sensitive applications.
IETF: RFC 8289: Controlled Delay Active Queue Management
Documents CoDel, an active queue management approach designed to control persistent queue delay, especially at consumer-edge bottlenecks.
Bufferbloat.net: Tests for Bufferbloat
Explains why useful bufferbloat tests measure latency during download and upload load instead of relying on an idle ping alone.
Bufferbloat.net: Bufferbloat FAQs
Provides practical interpretation and notes that a latency-under-load increase below roughly 15–25 ms is generally well controlled.

FAQ

Bufferbloat test questions

What is bufferbloat?: Bufferbloat is excessive delay caused when packets wait in an overfilled queue at a bottleneck. It often appears only while a download, upload, cloud sync, or other capacity-seeking transfer is active.
How does this browser bufferbloat test work?: The test first measures idle browser HTTP round-trip time. It then runs separate download and upload transfers while continuing small latency probes, producing retained median, p95 tail, jitter, discarded-sample, and failed-application-probe diagnostics.
What is a good bufferbloat result?: There is no universal letter-grade standard. SwiftSpeedTest labels a run A when the worst retained median rises no more than 15 ms, the worst p95 rises no more than 40 ms, loaded jitter is no more than 10 ms, and its stricter sample-quality boundaries are also met.
Does a poor result prove my router is the problem?: No. Queueing can occur on Wi-Fi, a router, modem, access link, VPN path, provider network, or remote test path. Compare Ethernet and Wi-Fi while keeping the device, endpoint, and time window fixed before isolating a layer.
Does this test measure packet loss?: It reports failed or timed-out browser HTTP application probes. Those misses are useful quality evidence, but they are not an ICMP, UDP, RTP, or network-layer packet-loss measurement.
How can I reduce latency under load?: First identify whether download or upload load is worse and repeat on Ethernet. Then pause uncontrolled background traffic, update the router, and evaluate documented smart queue management or active queue management support. Change one setting at a time and retest against the same endpoint.
How much data does the test use?: The page starts in full-capacity mode because a bufferbloat test must create load. Before starting, the tester shows a time- and byte-bounded maximum estimate and offers a low-data mode, although low-data mode may not fully saturate a fast connection.

Grade

Median rise

p95 tail rise

Loaded jitter

Discarded samples

HTTP probe misses

≤ 15 ms

≤ 40 ms

≤ 10 ms

< 5%

< 1%

≤ 30 ms

≤ 75 ms

≤ 20 ms

< 10%

< 2%

≤ 60 ms

≤ 150 ms

≤ 40 ms

< 20%

< 5%

≤ 120 ms

≤ 300 ms

≤ 80 ms

< 35%

< 10%

One or more metrics exceed every D boundary.

Capacity and responsiveness are different

Download rise: queueing while the connection receives sustained traffic.

Upload rise: queueing while backups, calls, or other traffic sends data.

p95 tail: the slower end of the successful probe series, preserved so a stable median does not hide repeated stalls.

HTTP misses: unsuccessful application probes that must not be relabeled as ICMP or UDP packet loss.

Isolate the queue before changing settings

1. Repeat once. Keep the endpoint, device, and connection type fixed to rule out a single noisy run.

2. Compare Ethernet and Wi-Fi. A large wireless-only change points toward the local radio, driver, or access point rather than proving an ISP issue.

3. Check the worse direction. Upload-only inflation often appears during cloud backup, camera sync, or large sends; download-only inflation needs a different controlled reproduction.

4. Evaluate documented queue management. If the router supports SQM, FQ-CoDel, CAKE, CoDel, or PIE, follow that device's instructions and retest after one change. Do not assume a label alone proves the right setting.

For intermittent problems without intentional load, use the internet stability test.

Methodology sources

IETF: RFC 7928: Characterization Guidelines for AQM

Describes bufferbloat as excessive unmanaged buffering that increases end-to-end delay and harms latency-sensitive applications.

IETF: RFC 8289: Controlled Delay Active Queue Management

Documents CoDel, an active queue management approach designed to control persistent queue delay, especially at consumer-edge bottlenecks.

Bufferbloat.net: Tests for Bufferbloat

Explains why useful bufferbloat tests measure latency during download and upload load instead of relying on an idle ping alone.

Bufferbloat.net: Bufferbloat FAQs

Provides practical interpretation and notes that a latency-under-load increase below roughly 15–25 ms is generally well controlled.

Bufferbloat test questions

What is bufferbloat?

Bufferbloat is excessive delay caused when packets wait in an overfilled queue at a bottleneck. It often appears only while a download, upload, cloud sync, or other capacity-seeking transfer is active.

How does this browser bufferbloat test work?

The test first measures idle browser HTTP round-trip time. It then runs separate download and upload transfers while continuing small latency probes, producing retained median, p95 tail, jitter, discarded-sample, and failed-application-probe diagnostics.

What is a good bufferbloat result?

There is no universal letter-grade standard. SwiftSpeedTest labels a run A when the worst retained median rises no more than 15 ms, the worst p95 rises no more than 40 ms, loaded jitter is no more than 10 ms, and its stricter sample-quality boundaries are also met.

Does a poor result prove my router is the problem?

No. Queueing can occur on Wi-Fi, a router, modem, access link, VPN path, provider network, or remote test path. Compare Ethernet and Wi-Fi while keeping the device, endpoint, and time window fixed before isolating a layer.

Does this test measure packet loss?

It reports failed or timed-out browser HTTP application probes. Those misses are useful quality evidence, but they are not an ICMP, UDP, RTP, or network-layer packet-loss measurement.

How can I reduce latency under load?

First identify whether download or upload load is worse and repeat on Ethernet. Then pause uncontrolled background traffic, update the router, and evaluate documented smart queue management or active queue management support. Change one setting at a time and retest against the same endpoint.

How much data does the test use?

The page starts in full-capacity mode because a bufferbloat test must create load. Before starting, the tester shows a time- and byte-bounded maximum estimate and offers a low-data mode, although low-data mode may not fully saturate a fast connection.