X hits on this document

PDF document

google-cluster-auctions.pdf - page 7 / 8

22 views

0 shares

0 downloads

0 comments

7 / 8

Engineering Teams

final prices and allocations

resource requests

preliminary prices

100

Resource Markets Front End

80

simulated auction mapping

Bidder Proxies

Utilization Percentile

60

40

Clock Auction Simulator

20

Fig. 5.

Bidder interaction with market and clock auction.

CPU Bids

CPU Offers

RAM Bids

RAM Offers

Disk Bids

Disk Offers

Market Price / Fixed Price

Auction

Median of γu

Mean of γu

% Settled

1

0.0092

0.0614

58.9%

2

0.0025

0.2078

88.2%

3

0.0009

0.0202

50.0%

2.0

1.5

1.0

CPU RAM Disk

Fig. 7.

Utilization percentiles of resources in settled transactions.

TABLE I BID PREMIUM STATISTICS.

0.5

to act on those costs autonomously. Without this market, the company would be forced to make centralized policy choices about the best placement of teams with imperfect knowledge of the engineering tradeoffs required.

C. Bidder Behavior

0.0

r1 r2 r3 r4 r5 r6 r7 r8 r9 r10 r11 r12 r13 r14 r15 r16 r17 r18 r19 r20 r21 r22 r23 r24 r25 r26 r27 r28 r29 r30 r31 r32 r33 r34

Cluster

Fig. 6.

Change in resource prices after auction.

bids and offers in three resource dimensions. This plot shows that most bids were for resources in underutilized clusters and most offers were for resources in overutilized clusters, which was the behavior strongly encouraged by the utilization- weighted reserve prices used to start the clock auction. It is also interesting to note the significant number of outliers, each representing resource needs for teams willing to pay a large premium.

It is worth noting that in those clusters with the highest market prices for resources we saw a number of large teams offer resources on the market to take advantage of the higher prices and move to less congested clusters. We also saw other teams that were willing to pay a significant price premium to continue growing in congested clusters even though resources were available at much lower cost elsewhere. This discrepancy shows the different premiums teams placed on relocation. There is an engineering cost to reconfiguring applications for different resource pools and the market economy allows teams

As the internal market economy has evolved we have noticed a number of distinct changes in bidder behavior. As users become more familiar with the market prices we have seen the reserve prices associated with bids move from closely tracking the former fixed price values to values much closer to the dynamic market prices. In particular, we define the premium between the bid price and the ultimate settled price as

γu =

π u x T u p u x T u p u

(5)

for each winning user, u W.

Table I shows the mean and median of this premium for all bids in the last three auctions. The fourth column shows the percentage of trades that were ultimately settled in that auction. In the earlier auctions bid prices were at times wildly divergent, but the median has decreased significantly over time. It is worth pointing out that the mean has been more variable as in some auctions a number of sellers will enter very low prices confident that there will be ample competition and that the final market price will be fair. Similarly, some bidders in earlier auctions would enter arbitrarily low bids in the expectation that these trades would be settled due to lack of competition and excess Google supply without reserve prices.

Document info
Document views22
Page views22
Page last viewedSun Dec 04 21:36:47 UTC 2016
Pages8
Paragraphs264
Words6974

Comments