Skip to content
Athrael.net logoAthrael.net

« Evaluating Narwhal

Mixed workload results

FieldValue
Evaluation dates17 to 18 September 2026
WorkloadFive request classes across nine consecutive load phases
Modelmoonshotai/Kimi-K3
FleetSix TP8 engines
FrameworksNarwhal; Dynamo Planner; Ray Serve LLM
Prefix-caching modesOff; On
Offers per configuration14,419
TTFT SLOAt or below 2.1 seconds
TPOT SLOAt or below 45 milliseconds

Serving versions and settings are on the serving configurations page.

Request outcomes and SLO delivery

FrameworkPrefix cachingOfferedCompletedFailedRefusedTTFT SLO deliveredTPOT SLO deliveredJoint latency SLO deliveredSLO-qualified
Dynamo PlannerOff14,4198,298 (57.55%)19 (0.13%)6,102 (42.32%)4,554 (31.58%)8,317 (57.68%)4,554 (31.58%)2,144 (14.87%)
Dynamo PlannerOn14,4198,131 (56.39%)11 (0.08%)6,277 (43.53%)6,027 (41.80%)8,142 (56.47%)6,027 (41.80%)3,135 (21.74%)
NarwhalOff14,41913,999 (97.09%)48 (0.33%)372 (2.58%)12,167 (84.38%)13,999 (97.09%)12,167 (84.38%)6,327 (43.88%)
NarwhalOn14,41913,973 (96.91%)51 (0.35%)395 (2.74%)12,427 (86.18%)13,973 (96.91%)12,427 (86.18%)6,579 (45.63%)
Ray Serve LLMOff14,41912,364 (85.75%)2,055 (14.25%)0 (0.00%)7,943 (55.09%)12,362 (85.73%)7,943 (55.09%)3,670 (25.45%)
Ray Serve LLMOn14,41913,569 (94.11%)850 (5.89%)0 (0.00%)13,462 (93.36%)10,825 (75.07%)10,767 (74.67%)5,353 (37.12%)

Failure stage

FrameworkPrefix cachingRequest classFailedBefore first output tokenAfter output started
Dynamo PlannerOffConversation History101
Dynamo PlannerOffDocument000
Dynamo PlannerOffLong Generation000
Dynamo PlannerOffShort Chat909
Dynamo PlannerOffStandard Chat909
Dynamo PlannerOnConversation History000
Dynamo PlannerOnDocument000
Dynamo PlannerOnLong Generation202
Dynamo PlannerOnShort Chat202
Dynamo PlannerOnStandard Chat707
NarwhalOffConversation History000
NarwhalOffDocument31310
NarwhalOffLong Generation330
NarwhalOffShort Chat11110
NarwhalOffStandard Chat330
NarwhalOnConversation History110
NarwhalOnDocument31310
NarwhalOnLong Generation770
NarwhalOnShort Chat880
NarwhalOnStandard Chat440
Ray Serve LLMOffConversation History34340
Ray Serve LLMOffDocument1,2951,2950
Ray Serve LLMOffLong Generation61610
Ray Serve LLMOffShort Chat4824820
Ray Serve LLMOffStandard Chat1831830
Ray Serve LLMOnConversation History770
Ray Serve LLMOnDocument4824820
Ray Serve LLMOnLong Generation39390
Ray Serve LLMOnShort Chat2392390
Ray Serve LLMOnStandard Chat83830

Recorded terminal signatures

FrameworkPrefix cachingOutcomeHTTP statusCauseRequests
Dynamo PlannerOffFailed20050019
Dynamo PlannerOffRefused529http 529 refusal6,102
Dynamo PlannerOnFailed20050011
Dynamo PlannerOnRefused529http 529 refusal6,277
NarwhalOffFailed504queue expired48
NarwhalOffRefused429server overloaded error372
NarwhalOnFailed504queue expired51
NarwhalOnRefused429server overloaded error395
Ray Serve LLMOffFailed503service unavailable2,055
Ray Serve LLMOnFailed503service unavailable850

Completion termination

FrameworkPrefix cachingNatural completionsOutput-cap completions
Dynamo PlannerOff4,404 (53.07%)3,894 (46.93%)
Dynamo PlannerOn4,267 (52.48%)3,864 (47.52%)
NarwhalOff7,444 (53.18%)6,555 (46.82%)
NarwhalOn7,470 (53.46%)6,503 (46.54%)
Ray Serve LLMOff6,443 (52.11%)5,921 (47.89%)
Ray Serve LLMOn7,185 (52.95%)6,384 (47.05%)

Answer quality

FrameworkPrefix cachingScored offersAll-offer token F1Completed-only token F1SLO-delivered token F1
Dynamo PlannerOff14,41917.0010%29.5417%7.0999%
Dynamo PlannerOn14,41916.1254%28.5958%10.6774%
NarwhalOff14,41928.5651%29.4221%21.8589%
NarwhalOn14,41928.3793%29.2851%22.6357%
Ray Serve LLMOff14,41924.2343%28.2622%11.9505%
Ray Serve LLMOn14,41927.5034%29.2263%18.1647%

Completed-request latency

FrameworkPrefix cachingTTFT p50 (s)TTFT p95 (s)TTFT p99 (s)TPOT p50 (ms)TPOT p95 (ms)TPOT p99 (ms)
Dynamo PlannerOff1.9543.9605.01835.91636.64836.987
Dynamo PlannerOn1.4053.5204.66736.04036.68737.091
NarwhalOff0.57611.20413.46032.70136.18836.660
NarwhalOn0.38411.26814.20332.76036.25636.772
Ray Serve LLMOff0.89316.89818.62323.50826.12528.425
Ray Serve LLMOn0.4891.1831.91726.22749.74750.313

Completed-request E2E latency

FrameworkPrefix cachingE2E p50 (s)E2E p95 (s)E2E p99 (s)
Dynamo PlannerOff10.58460.444113.011
Dynamo PlannerOn10.35162.339114.572
NarwhalOff8.74255.266109.099
NarwhalOn8.45054.768108.823
Ray Serve LLMOff11.32547.78687.713
Ray Serve LLMOn7.73852.81295.383

Outcomes by request class

FrameworkPrefix cachingRequest classOfferedCompletedFailedRefusedSLO-qualified
Dynamo PlannerOffConversation History474311 (65.61%)1 (0.21%)162 (34.18%)89 (18.78%)
Dynamo PlannerOffDocument6,3103,539 (56.09%)0 (0.00%)2,771 (43.91%)1,141 (18.08%)
Dynamo PlannerOffLong Generation550252 (45.82%)0 (0.00%)298 (54.18%)45 (8.18%)
Dynamo PlannerOffShort Chat4,9562,887 (58.25%)9 (0.18%)2,060 (41.57%)490 (9.89%)
Dynamo PlannerOffStandard Chat2,1291,309 (61.48%)9 (0.42%)811 (38.09%)379 (17.80%)
Dynamo PlannerOnConversation History474290 (61.18%)0 (0.00%)184 (38.82%)85 (17.93%)
Dynamo PlannerOnDocument6,3103,259 (51.65%)0 (0.00%)3,051 (48.35%)1,966 (31.16%)
Dynamo PlannerOnLong Generation550258 (46.91%)2 (0.36%)290 (52.73%)59 (10.73%)
Dynamo PlannerOnShort Chat4,9562,934 (59.20%)2 (0.04%)2,020 (40.76%)579 (11.68%)
Dynamo PlannerOnStandard Chat2,1291,390 (65.29%)7 (0.33%)732 (34.38%)446 (20.95%)
NarwhalOffConversation History474468 (98.73%)0 (0.00%)6 (1.27%)184 (38.82%)
NarwhalOffDocument6,3106,065 (96.12%)31 (0.49%)214 (3.39%)4,006 (63.49%)
NarwhalOffLong Generation550531 (96.55%)3 (0.55%)16 (2.91%)171 (31.09%)
NarwhalOffShort Chat4,9564,838 (97.62%)11 (0.22%)107 (2.16%)1,127 (22.74%)
NarwhalOffStandard Chat2,1292,097 (98.50%)3 (0.14%)29 (1.36%)839 (39.41%)
NarwhalOnConversation History474469 (98.95%)1 (0.21%)4 (0.84%)187 (39.45%)
NarwhalOnDocument6,3106,058 (96.01%)31 (0.49%)221 (3.50%)4,207 (66.67%)
NarwhalOnLong Generation550526 (95.64%)7 (1.27%)17 (3.09%)170 (30.91%)
NarwhalOnShort Chat4,9564,830 (97.46%)8 (0.16%)118 (2.38%)1,167 (23.55%)
NarwhalOnStandard Chat2,1292,090 (98.17%)4 (0.19%)35 (1.64%)848 (39.83%)
Ray Serve LLMOffConversation History474440 (92.83%)34 (7.17%)0 (0.00%)167 (35.23%)
Ray Serve LLMOffDocument6,3105,015 (79.48%)1,295 (20.52%)0 (0.00%)1,774 (28.11%)
Ray Serve LLMOffLong Generation550489 (88.91%)61 (11.09%)0 (0.00%)144 (26.18%)
Ray Serve LLMOffShort Chat4,9564,474 (90.27%)482 (9.73%)0 (0.00%)909 (18.34%)
Ray Serve LLMOffStandard Chat2,1291,946 (91.40%)183 (8.60%)0 (0.00%)676 (31.75%)
Ray Serve LLMOnConversation History474467 (98.52%)7 (1.48%)0 (0.00%)179 (37.76%)
Ray Serve LLMOnDocument6,3105,828 (92.36%)482 (7.64%)0 (0.00%)3,159 (50.06%)
Ray Serve LLMOnLong Generation550511 (92.91%)39 (7.09%)0 (0.00%)187 (34.00%)
Ray Serve LLMOnShort Chat4,9564,717 (95.18%)239 (4.82%)0 (0.00%)1,043 (21.05%)
Ray Serve LLMOnStandard Chat2,1292,046 (96.10%)83 (3.90%)0 (0.00%)785 (36.87%)

Answer quality by request class

FrameworkPrefix cachingRequest classScored offersAll-offer token F1Completed-only token F1SLO-delivered token F1
Dynamo PlannerOffConversation History47413.7066%20.8904%7.3886%
Dynamo PlannerOffDocument6,31027.3484%48.7620%10.3557%
Dynamo PlannerOffLong Generation5508.4483%18.4387%3.3382%
Dynamo PlannerOffShort Chat4,9566.0550%10.3943%3.1610%
Dynamo PlannerOffStandard Chat2,12914.7567%24.0008%7.5268%
Dynamo PlannerOnConversation History47411.9914%19.5998%7.0334%
Dynamo PlannerOnDocument6,31024.9341%48.2768%17.6084%
Dynamo PlannerOnLong Generation5509.0745%19.3448%3.7917%
Dynamo PlannerOnShort Chat4,9566.2247%10.5144%3.7443%
Dynamo PlannerOnStandard Chat2,12915.8074%24.2115%8.8645%
NarwhalOffConversation History47419.8877%20.1427%15.4645%
NarwhalOffDocument6,31046.3692%48.2423%36.1500%
NarwhalOffLong Generation55018.0898%18.7371%12.0006%
NarwhalOffShort Chat4,95610.3636%10.6164%7.6363%
NarwhalOffStandard Chat2,12922.8053%23.1533%16.5807%
NarwhalOnConversation History47419.9770%20.1900%15.1476%
NarwhalOnDocument6,31046.0606%47.9766%37.8770%
NarwhalOnLong Generation55017.4523%18.2486%11.5444%
NarwhalOnShort Chat4,95610.3374%10.6071%7.7393%
NarwhalOnStandard Chat2,12922.6671%23.0901%16.6720%
Ray Serve LLMOffConversation History47418.7679%20.2182%13.5435%
Ray Serve LLMOffDocument6,31037.9926%47.8033%16.0383%
Ray Serve LLMOffLong Generation55016.2249%18.2488%9.9025%
Ray Serve LLMOffShort Chat4,9569.3828%10.3936%6.1609%
Ray Serve LLMOffStandard Chat2,12921.3150%23.3194%13.4870%
Ray Serve LLMOnConversation History47419.8272%20.1244%14.6564%
Ray Serve LLMOnDocument6,31044.4116%48.0846%28.3475%
Ray Serve LLMOnLong Generation55017.0242%18.3235%13.0031%
Ray Serve LLMOnShort Chat4,95610.0711%10.5813%7.2067%
Ray Serve LLMOnStandard Chat2,12922.3865%23.2947%15.6075%

Queue telemetry

FrameworkPrefix cachingQueue scopeQueue wait p50 (s)Queue wait p95 (s)Queue wait p99 (s)Queued requests p50Queued requests p95Queued requests peak
Dynamo PlannerOffDynamo request plane and frontend HTTP queue0.0000870.0002280.0002505.0012.0018.00
Dynamo PlannerOnDynamo request plane and frontend HTTP queue0.0000890.0002190.0002384.008.0016.00
NarwhalOffNarwhal admission queue0.0003639.30546614.3550200.000.004.00
NarwhalOnNarwhal admission queue0.0002669.33653315.2160550.000.004.00

Definitions